Lenovo, 그린 전력을 투입해 토큰을 산출하는 방식으로 '토큰 팩토리' 개념을 재정의하며, 연간 2,000조 개의 토큰 생산량과 70% 비용 절감을 주장했다.

Output 2,000 trillion Tokens a year, with costs cut by 70%.

Buying the chips is affordable, but using AI is not. That may sound far-fetched, yet as AI spreads across enterprises, many companies have found themselves in exactly this bind.

AI's growth over the past two years can only be described as explosive. According to the National Data Administration, China's average daily Token consumption rose from 100 billion in early 2024 to 100 trillion by the end of 2025, passed 140 trillion in March 2026, and grew roughly a thousandfold in two years. The China Academy of Information and Communications Technology reported that the relevant figure had surpassed 500 trillion by June 2026.

Surging demand quickly reached the compute supply side. As models evolve and agents take on increasingly complex tasks, Token demand has soared, and computing infrastructure has expanded to match. IDC expects global server market revenue to reach about US$685.1 billion in 2026, with AI infrastructure investment as the main driver.

Yet even as compute supply keeps growing, enterprises are becoming more cautious about "burning Tokens."

According to a report by the *Economic Observer*, a mid-sized bank spent nearly 9 million yuan on Tokens in the first half of the year, equal to 8% of its previous year's IT budget. In early the year, rising Token consumption was seen as a positive sign that AI was entering core business operations. By the end of June, however, the bank's management asked for a review of the cost-effectiveness of these investments.

A McKinsey survey likewise found that about one in five respondents said their organizations had restricted AI use because of AI operating costs.

Enterprises have begun measuring the cost of AI against its actual output.

The reasons are not hard to understand. Complex agent tasks can run for hours, and repeated reasoning, model calls, and task retries add up. Self-built compute must be provisioned for peak business loads, and depreciation, electricity, and maintenance costs continue even when the equipment sits idle.

Worse still, the spending does not always go where it is needed. Simple tasks may run on expensive models, while sensitive operations must stay in private environments. Compute is hard to allocate where it belongs.

As a result, enabling enterprises to use AI well and affordably remains an urgent problem.

On the windswept plateau of Zhangjiakou in Hebei, a facility is now rising: a "hybrid Token factory" with a planned 15,000P supernode cluster and an annual capacity of 2,000 trillion Tokens. On November 10, Lenovo Group officially unveiled the "Lenovo Wanquan Hybrid Token Factory" brand locally.

Rather than simply delivering servers or selling computing resources, the factory plans to organize compute and models from different sources to continuously produce more economical and efficient Token services.

Its production logic is strikingly direct: green power in, Tokens out.

The era of brute-force chip stacking is giving way to an era of industrialized, cost-effective Token mass production. The underlying rules of computing infrastructure may be about to change overnight.

**Green power in, Tokens out**

On the same day, Lenovo Group and the Zhangjiakou municipal government signed a cooperation agreement to advance the first hybrid Token parent factory in Zhangjiakou.

According to the latest plan, the project's first phase calls for a 15,000P heterogeneous compute cluster and a roughly 100 MW intelligent computing center. Once completed, the facility is expected to produce 2,000 trillion Tokens annually, with a peak output of 200 million Tokens per second.

Dai Wei, Senior Vice President of Lenovo Group and General Manager of its China Solutions and Services Business Group, said on site that the first phase is scheduled for 2027, and that by 2030 the scale is planned to expand by 10 times or more.

What does an annual capacity of 2,000 trillion Tokens mean? On average, that is about 5.5 trillion Tokens produced per day. Even this simple conversion shows the scale involved.

Yet the factory's name may be even more intriguing than its size.

The concept of a Token factory is not new. Existing industry efforts have taken different approaches: some start from chip and server energy efficiency, some operate model platforms, and others specialize in intelligent computing center operations. Each technical route has its focus, but when enterprises actually use AI, they often face a combination of different computing needs.

What sets Lenovo's Token factory apart is captured in one word: "hybrid." At its core, "hybrid" means letting compute supply follow business demand.

Enterprises need stable, controllable private compute, but they also face sudden demand for public calls. Different tasks also vary widely in their requirements for model capability, response speed, and cost.

Simply expanding compute capacity, or optimizing just one stage of Token production, cannot solve all of these problems on its own.

Lenovo's proposed "hybrid supply" acts as a hub that reorganizes previously scattered resources. Private and public Token services are combined, with private resources as the primary base; device, edge, and cloud operate in coordination; and different models can be flexibly selected according to task requirements.

In this way, compute supply can follow business needs rather than being determined entirely by hardware configuration.

The ideal is for Token supply to work like electricity. Enterprises connect on demand and receive stable, economical service, while the equipment behind it, the models invoked, and the scheduling of resources are all handled by a unified supply system.

Of course, making different resources work together within a single system requires more than models and servers. Computing equipment, data center operations, energy supply, and even model scheduling all need to be considered.

How can this be ensured?

Lenovo has set its sights on a "wind hub" close to Beijing: Zhangjiakou in Hebei.

It is called a wind hub first because of its rare energy endowment.

Zhangjiakou is the national leader in wind and solar green power and a benchmark dual-node city in China's "Six Networks" (computing power network plus new-type power grid). By the end of 2025, its renewable energy installed capacity had exceeded 45 million kilowatts, accounting for nearly 90% of the city's total installed power capacity.

Second, the city already has a considerable computing foundation.

Zhangjiakou is an important host for the Beijing-Tianjin-Hebei hub of the national "East Data, West Computing" project, and the local government is actively providing policy support and industrial guidance. So far, the city has built more than 50 data centers, with over 1.07 million standard racks in operation.

Zhangjiakou's abundant wind and solar resources, cool climate, and competitive electricity prices, combined with its location just over 100 kilometers from Beijing and the lower latency that comes with it, can help reduce computing operating costs and meet AI inference demand across the Beijing-Tianjin-Hebei region.

Policy is also pushing the computing industry toward more efficient supply. The "15th Five-Year Plan Outline," released this year, calls for promoting the scaled, intensive, green, and inclusive development of computing power, fostering coordination between green electricity and computing, and lowering computing costs for small and medium-sized enterprises.

Working with local authorities, Lenovo has thus explored a more "end-to-end" industrial model: green power in, Tokens out.

It may sound like an industrial assembly line. But unlike a traditional factory, a Token production line must not only have stable energy and equipment; it must also constantly adjust its production methods to meet widely varying task requirements.

So how can compute in different locations and models with different capabilities truly work together to serve enterprises?

**Putting every unit of compute to best use**

Consider a manufacturing company that uses AI extensively. It might face this scenario:

Production systems must run stably, so sensitive data cannot leave the private environment. Yet office assistants and customer-service agents may suddenly receive a flood of requests during peak business hours.

If it builds all compute in-house, the company must maintain equipment for short-lived peaks, at enormous cost. If it relies entirely on purchased public services, it still runs into data boundaries, ongoing expenses, and model-adaptation challenges.

Moreover, simple document organization and complex professional reasoning should not run on the same expensive configuration. Clearly, different tasks require different ways of supplying compute.

Following this logic, Lenovo proposed its hybrid supply model, which has three dimensions:

- **Hybrid supply sources**: privately produced Tokens complement licensed public services.

- **Coordinated device, edge, and cloud**: lightweight tasks stay on devices, latency-sensitive tasks run nearby, and elastic demand is matched to the cloud.

- **Hybrid model ecosystem**: open-weight or proprietary closed-source models are selected according to quality and cost requirements.

More choices, however, also mean more complex scheduling.

Factors such as whether a task involves sensitive data, how powerful a model it needs, how long it can wait, and how private and public resources should be allocated all have to be weighed together. Allocation between private and public resources is especially tricky because the two types of workload have opposite characteristics. Resources that run continuously inside the enterprise must be stable and keep data within bounds, while those that rise and fall with external calls must be economical and scale with peaks.

To address this, Lenovo designed the **Token Engine**, which improves resource matching through dynamic co-location of models and algorithms, task-level resource pool allocation, and intelligent model routing.

For scheduling to truly span private and public resources, a unified management system is also needed.

Lenovo's existing **xCloud (Lenovo Intelligent Cloud)** fills this role. It can centrally manage, schedule, and intelligently operate private and public cloud resources, enabling computing resources scattered across different environments to work together.

Building on this, xCloud includes the **Token Hub**. It focuses specifically on AI inference scenarios, using model routing and dynamic allocation of inference resource pools to match each request with the appropriate model and compute, based on task requirements, model quality, data boundaries, response latency, and cost.

As a result, previously scattered resources can be flexibly combined according to business needs. Scheduling can even reach down to individual **chips and compute nodes**.

Yao Zhenyu, Vice President of Lenovo Group and General Manager of Global Infrastructure and Intelligent Cloud, gave an example: lightweight models can run on smaller compute nodes or edge resources without occupying large clusters.

In other words, scheduling is not just about finding idle machines. It can also ensure that tasks **use the right model and chip**, so computing resources are fully utilized and every unit of compute is put to its best use.

In addition, to ensure data security and compliance, Lenovo has set up barriers in three dimensions:

- **Model security**: all models must pass security testing and certification and enter a whitelist before enterprises can use them on the Token Hub.

- **Content security**: Lenovo's **Token Trust** capability actively intercepts sensitive information along the transmission path, and sets data security and usage permissions as boundaries that resource scheduling must observe.

- **Privacy computing and TEE (Trusted Execution Environment)**: Lenovo supports processing private and sensitive data in isolated trusted environments, where the platform cannot read the content and only authorized users can access it. This is what the industry calls "usable but invisible."

No amount of fine-tuned scheduling, however, can erase the cost of the equipment and data center themselves. Lenovo therefore proposed **three levels of optimization**:

- **Level one: improving compute equipment efficiency.** Within its **Wentian heterogeneous intelligent computing solution**, Lenovo combines supernodes, a multi-chip, multi-model training and inference engine, and optimization of cluster communication and operators to improve the coordination of heterogeneous compute resources. Lenovo says the solution can deliver more than 30% improvement in compute efficiency. For large clusters, reducing some waiting and idle time can release considerable computing capacity.

- **Level two: controlling the full lifecycle cost of intelligent computing centers.** Beyond servers, power supply, cooling, construction, and operations are equally expensive. Lenovo optimizes data center construction and daily operations together through modular prefabrication, integrated full liquid cooling, and green DC power intelligent control. Lenovo also plans to build a computing equipment recycling and reuse line in Zhangjiakou to extend the useful life of existing equipment, and the supporting intelligent computing center targets a PUE below 1.1.

- **Level three: Token efficiency optimization for business scenarios.** Through model routing and dynamic allocation of resource pools, simple tasks need not occupy high-cost models, and sudden demand can be matched with appropriate compute.

In this way, hardware, data centers, and the Token Hub each optimize one link, and the resources and costs saved at each stage ultimately add up to greater efficiency for enterprises using AI. As Lenovo puts its core message, the goal is to "release every unit of computing efficiency."

To be accountable to enterprise customers, Lenovo also validated the approach with its own business. Yao Zhenyu revealed that through internal AI practice, the company has achieved about 70% overall cost optimization.

Lenovo has now set overall targets for the hybrid Token factory, including a 70% reduction in the comprehensive cost per million Tokens, a fivefold increase in Token output per second, and a Token yield rate above 99.9%.

Of course, measuring the efficiency of a Token factory cannot rely on cost and throughput alone.

No matter how large the capacity or how low the cost, response speed and model quality must be preserved. Sensitive enterprise data in particular cannot be sacrificed in the name of cost reduction. And the energy consumption of large intelligent computing centers is a key factor determining long-term operating costs.

Therefore, in addition to **overall cost and Token throughput**, Lenovo also considers **task-completion accuracy, green and sustainable operation, and end-to-end security** as the five core competitive dimensions of the Token factory.

In Lenovo's summary, these are "many, fast, good, economical, and green."

"Good" deserves particular attention. When an agent completes a task, one system may deliver a qualified result on the first try, while another requires repeated revisions and reasoning. Even if the latter consumes more Tokens, it does not necessarily create more value.

Lenovo offers a vivid concept for this: **effective Tokens**.

For enterprises, Token output and the unit price per million Tokens matter, but without task-completion quality, even impressive performance metrics struggle to show how much real value AI has created.

**Turning a factory into a method**

"Each part may be right, yet the whole goes wrong." Yao Zhenyu said he has seen many such cases.

When an enterprise IT system fails, the hardware, database, and platform vendors can each prove that their part is working, yet the root cause of the failure remains elusive.

The more links an AI system involves, the more pronounced such dilemmas become.

The threshold a Token factory must cross is therefore precisely to **reorganize scattered capabilities**.

This is the opportunity Lenovo targeted from the start of building its Token factory.

As a hardware veteran with business experience across many technical links, Lenovo summarizes its enterprise AI business as "three-layer hybrid power":

- **Foundation layer**: the **Wentian AI infrastructure**, covering servers, storage, networking, and heterogeneous intelligent computing platforms.

- **Middle layer**: the **Wanquan Hybrid Token Factory** unveiled this time, which converts computing resources into sustainable Token services.

- **Upper layer**: **agent solutions and services** such as **Qingtian** and **Baiying**, which connect directly to enterprises' practical applications.

These three layers span compute production, Token supply, and business use, pointing toward a hub role that helps more enterprises shift from "+AI" to "AI+."

"+AI" means introducing AI tools into existing business processes to solve problems in local links. "AI+" means having AI participate in core operations such as R&D, production, and supply chains, driving the restructuring of production processes and organizational methods.

The more complex the tasks AI handles and the deeper it goes into business operations, the more enterprises depend on the stability, economics, and flexibility of the underlying compute.

Lenovo's strength lies in its business capabilities spanning both infrastructure and applications. It can participate in optimizing equipment, energy, and compute efficiency, while also gaining feedback from real business through model services and agent applications.

Connecting the two ends of compute production and real business places Lenovo closer to the final output along the AI infrastructure value chain.

Having full capabilities is one thing, however; letting collaboration happen naturally is another. To turn cross-link optimization into a **replicable method**, a mature testing ground is needed.

Zhangjiakou therefore serves not only as the first site of the Lenovo Wanquan Hybrid Token Factory but also as a validation ground. Lenovo calls it the "first hybrid Token parent factory."

Lenovo plans to use its own real AI business workloads to test the collaborative efficiency of different compute resources and models, and to continuously optimize resource scheduling, cost control, and security operations. It will then extend mature technologies and operational experience to different industries.

Liu Jun, Executive Vice President of Lenovo Group and President of Lenovo China, noted that the company already has more than 100 agent applications internally, which provide a constantly changing set of usage scenarios.

The "parent factory" is not meant to be a data center that can be copied one-to-one. Instead, it aims to summarize a **method for making different resources work together**.

Consider the AI short-drama project with Hunan Broadcasting.

Dai Wei said Hunan Broadcasting's initial need was for a **content-generation agent**. As AI creation moved into routine production, **Token cost became a new pressure**. The two sides therefore extended their cooperation to the infrastructure layer, planning a hybrid Token factory in Changsha dedicated to the film and short-drama industry.

The story begins with the short-drama production process.

AI film and TV production involves scripts, storyboards, images, video, rendering, and other stages. The costs of different models vary widely, and project schedules create significant peaks and troughs in compute demand. Proprietary IP and copyrighted materials also raise data-security concerns, so they cannot be handed to external services without restriction.

Xiao Wenqiang, Investment Director of the Strategic Investment Department at Mango Media, offered a judgment on site: **the way compute is supplied must serve the way content is produced.**

Accordingly, the Hunan project plans to match computing resources to different production stages through technologies such as full-modality Token hybrid routing and multi-chip, multi-model inference acceleration.

According to the plan, the project aims to **reduce production costs by 50%, shorten production cycles by 60%, produce more than 1,200 hours of finished content per month, and aggregate more than 10,000 OPC (one-person company) content creators**.

1,200 hours equals an average of roughly 40 hours of finished content per day, which requires a large number of parallel tasks. Creation at this scale places far higher demands on the continuous supply of Tokens than scattered individual calls.

Similar tensions appear beyond the film and TV industry.

Education is one example. A Lenovo project with Shanghai Jiansheng General Education covers more than 600 primary and secondary schools. Daily teaching, collective lesson preparation, and exam grading create very different compute demands, and during winter and summer breaks, the load drops noticeably.

Unlike film production, however, the education system also involves large amounts of minors' information, with stricter requirements for **data security and service continuity**. Compute cannot be permanently provisioned for peak loads, nor can sensitive operations be moved to public environments simply to cut costs.

The challenge of hybrid supply, therefore, is not only adjusting how much resource to provide, but also finding the right balance between business patterns and data boundaries.

Whether demand is matched sensibly will ultimately be proven through long-term operations. The Gansu Qingyang Suihong Huachuang intelligent computing center, which Lenovo has disclosed, offers a reference from the operational side.

As the first hybrid Token operations case Lenovo has listed, the project reduces overall usage costs by about 30%. A single model processes more than 2 million calls and 2 billion Tokens per day on average, anchoring the metrics firmly in daily calls and actual costs.

The business workloads of these cases each have their own rhythm, and their resource constraints differ. For Token factories to reach industries across the economy, a single fixed hardware configuration cannot solve every problem.

What should be standardized is the underlying technology, scheduling, and operational methods. What should be differentiated is the supply approach tailored to specific industries.

This is the purpose behind Lenovo's parent factory. Zhangjiakou is responsible for validating the method, and other regions can combine it according to their energy conditions and industrial needs.

Once the method is validated, Lenovo has the opportunity to turn its business capabilities into a **replicable Token supply system**, becoming an **organizer of cross-regional and cross-industry Token supply**.

Yet the role of organizer means responsibility does not end at **delivery**.

When selling servers, performance and price have clear benchmarks. But once Token services enter production workflows, they must withstand long-term tests of business fluctuations, model adaptation, data security, and cost control.

The value of AI infrastructure is thus extending from delivering computing power to delivering real productivity.

For enterprises, **stable delivery of production tasks and controllable long-term costs** are the ultimate measures of AI infrastructure's value.

A similar shift can be traced back to the Industrial Revolution.

More than two centuries ago, the steam engine provided sustained and powerful mechanical energy. Factories gradually broke free from the constraints of waterways and geography, allowing more machines and workers to concentrate production.

The lesson of industrial production was not only greater power, but also that **energy, machines, and production processes were organized into an efficiently operating system**.

AI is now crossing a similar threshold.

Chip performance keeps climbing and model capabilities advance rapidly, yet between the compute in a data center and efficiency gains on the production floor lie many steps, including model deployment, resource scheduling, and business adaptation.

The key to the next round of AI infrastructure competition lies in turning **enormous computing capacity** into **sustained productivity**.

The more complex the technology, the less its users should be burdened by that complexity.

Heavy tasks such as model deployment, resource scheduling, and long-term operations should be handled by mature infrastructure, leaving enterprises free to focus their energy on products, production, and services.

Only then will AI's value truly return to solving problems.

And the AI productivity serving every industry will have a chance to be fully unleashed in everyday operations.