From Computing Power to Value: Reconstruction of Infrastructure and New Engine for Industrial Growth in the AI Era | 2026 AI Partner · Beijing Yizhuang AI+ Industry Conference

How does the token economy reshape the AI industrial chain? From chips to intelligent computing centers, from model services to terminal applications, tokens are becoming the pricing unit that runs through the entire chain. As the demand for inference computing power surpasses that for training, the role of intelligent computing centers is changing from a computing power warehouse to a token factory, and the door to a trillion - level market has just opened.

Tokens are becoming the new unit of productive forces in the AI era. Song Chen pointed out that as agents become the new interaction entrance, the token consumption for a single task has soared from a few thousand to the million - level, and China's token call volume already accounts for 61% of the global total. In this trend, the focus of competition in AI infrastructure is shifting from "whose model has more parameters" to "who can support the circulation of massive tokens at a lower cost". Yingbo Digital is positioned as the full - stack builder of the "token factory" - it doesn't develop models or sell applications, but provides a one - stop solution for intelligent computing centers from planning and design to the delivery of integrated hardware and software.

The following is the full text of the speech, edited by 36Kr:

Song Chen | Deputy General Manager of Beijing Yingbo Digital Technology Co., Ltd.

Distinguished guests, good afternoon! I'm Song Chen from Yingbo Digital. I'm very honored to be invited to the 2026 AI Partner Conference hosted by 36Kr to discuss the future of AI infrastructure with all of you here.

The theme of my speech today is - From computing power to value: The reconstruction of infrastructure and the new engine of industrial growth in the AI era.

In the following time, I'd like to share three core topics with you: First, how the token economy is reshaping the value anchor of the AI industry; Second, how the value chain from electricity to computing power to tokens is reconstructing the industrial growth path; Third, and the one I want to focus on most today - Yingbo Digital's thinking and practice in this round of industrial transformation as a full - stack intelligent computing infrastructure builder.

Okay, let's first enter the first part - the token economy. Why do I put tokens at the forefront of the entire speech? In my opinion, tokens are becoming the most basic value unit in the AI era. Only by understanding tokens can we understand the underlying logic of today's AI industry.

The explosion of the token economy is not driven by a single factor, but the result of resonance in three dimensions.

The first dimension is the qualitative change in application scenarios. As agents become the new interaction entrance, the token consumption for a single task has increased from a few thousand in the past to hundreds of thousands or even millions.

The second dimension is the establishment of a commercial closed - loop. Since 2025, with the implementation of more and more applications, especially the popularity of Agent applications such as Lobster this year, AI has truly entered the workflow and gradually formed a commercial closed - loop.

The third dimension is the setting of national strategy. Intelligent computing power has been clearly included in the new infrastructure project, which means that at the policy level, the foundation of the token economy has received the highest - level support.

Tokens are becoming the new unit of productive forces in the AI era.

The rise of the token economy has triggered a top - down reconstruction of the value of the industrial chain.

Let's take a look at this value chain diagram. At the chip layer, the specialization of GPUs and ASICs is accelerating, and the token computing density has become the core indicator. At the intelligent computing center layer, the biggest change we see is the transformation from a "computing power warehouse" to a "token production factory". At the model layer, the MaaS model has turned "technology" into a "commodity", and charging by tokens has become the standard. At the application layer, agents have become the new interaction entrance, evolving from "tools" to "productive forces".

What runs through these four levels? It's tokens. As a pricing unit, tokens run through from chip design, to intelligent computing centers, to model services, and then to terminal applications. A complete value chain is thus formed.

Where is Yingbo Digital positioned in this value chain? We focus on the intelligent computing center layer - that is, the "factory" link of token production. We provide full - stack capabilities for this factory from planning and design to the delivery of integrated hardware and software.

In this round of transformation, there is a very critical structural change - that is, the shift from "training - first" to "inference - first".

From this trend chart, we can see that in 2024, training computing power accounted for 60%, and it is expected to drop to 35% by 2027, while inference computing power will rise from 40% to 65%. This is not a simple trade - off, but a fundamental change in the computing paradigm.

Training pursues large - scale parallelism and throughput, while inference pursues low latency, high concurrency, and low cost per token. This directly determines that the design logic of intelligent computing centers is completely different - from the perspective of resource characteristics, training requires high - memory batch processing, while inference requires elastic scaling and real - time response. For those of us in the infrastructure field, this means that the entire technical architecture needs to be re - thought.

IDC's prediction shows that the scale of inference demand is expected to reach 5 to 10 times that of the training stage. In 2026, inference computing will account for more than 70% of the total computing demand for generative AI.

What's more worthy of attention is the increase in computing power demand brought about by multi - modal fusion - the context sequence length has increased from the thousand - level to the million - level, which means that the computing power consumption for a single request has increased exponentially.

These numbers tell us one thing: the inference era has arrived, and the infrastructure must be prepared for this new era.

Market data further verifies this. In 2025, the global intelligent computing center market reached $185 billion, and it is expected to exceed $550 billion by 2029. The growth rate of the Chinese market is even more rapid. In 2025, the market size of intelligent computing centers in China reached 135.6 billion yuan, and the total scale of intelligent computing power reached 1.59 billion P.

This is a trillion - level market with explosive growth, and we are at the starting point of this growth curve.

After understanding the underlying logic of the token economy, let's look at the second part - the new path of industrial growth. I'd like to put forward a core view: "Single - token cost" is becoming the new competitive benchmark for AI infrastructure.

We break down the value chain into three layers. The bottom layer is the power layer - electricity is becoming the "oil of the computing power era" and is the basic constraint for token production. "Computing - power and electricity coordination" has been included in the new infrastructure project, which is a national - level strategic layout.

The middle layer is the computing power layer - an intelligent computing center is essentially a token production factory. China has built 42 clusters with ten - thousand - card levels, and the intelligent computing power has reached 1.59 billion P. Computing power is changing from a cost center to the core of pricing power.

The top layer is the token layer - tokens are tradable "intelligent currencies". China's token call volume has surpassed that of the United States and accounts for 61% of the global total. This marks that we are shifting from a major exporter of manufacturing products to a major exporter of AI productive forces.

The business model is also undergoing a profound evolution. I summarize this evolution into three eras.

The 1.0 era is "selling hardware" - selling by GPU cards and servers, and recognizing revenue at one time. The 2.0 era is "selling resources" - renting computing power by machine - months and card - hours, similar to the "selling water and electricity" model.

Now, we are entering the 3.0 era - "selling intelligence". Charging by tokens, with hierarchical pricing, and outputting intelligent services themselves. This is not just a simple change in the billing method, but a leap in the entire industrial value proposition.

The latest market signals are very clear: In 2026, global cloud providers collectively raised prices by 30% to 100%, and the daily token call volume of domestic large models has exceeded 180 trillion. The seller's market has arrived, and computing power has changed from a cost center to the core of pricing power.

Currently, the curtain of the intelligent economy has just been opened, and token commercialization is still in its early stage. This is the best window period for layout.

In the previous two parts, we talked about the logic and trends. Now, let's enter the part I most want to share with you today - Yingbo Digital's practice. Who are we? We are not just suppliers of "production equipment", but the designers, builders, and formulators of operation and maintenance standards for the entire token factory.

This is our full - stack intelligent computing center solution, covering the entire life cycle of intelligent computing center construction from the underlying technical architecture to the upper - layer business operation.

At the cluster planning and construction level, we provide a standard four - plane networking solution - the management, storage, computing, and BMC networks are independent, realizing efficient isolation and scheduling of network traffic. This is the foundation for the stable operation of ten - thousand - card - level clusters.

At the integrated hardware and software delivery level, we provide a complete integration solution for GPU hosts, networks, and storage, and are equipped with the Yingbo Cloud GPU PaaS platform, which greatly reduces the integration difficulty and online cycle for customers.

At the flexible billing model level, we support four models: pay - as - you - go, annual or monthly packages, off - peak preemption, and timed reservation, meeting the diverse usage scenarios of customers - large - model training requires the stability of annual or monthly packages, digital marketing requires the flexibility of on - demand, video generation is suitable for the cost optimization of the spot model, and academic research and teaching training can be flexibly combined according to the cycle.

At the account system level, we have designed a sub - account and alliance account system, supporting resource allocation for multiple organizations and refined cost accounting, which is crucial for large - enterprise customers, research teams, and computing power alliance operators.

In terms of technical architecture, we have two core innovations. The first is the K8S Native architecture - giving customers bare - metal - level control while maintaining SaaS - level ease of use. The second is the DICP dynamic isolation control plane - each user has an independent API Server and CRD space, fundamentally ensuring the security and isolation in a multi - tenant environment.

In a nutshell: With full - stack technical capabilities and rich project practices, Yingbo Digital provides a one - stop intelligent computing center solution from planning, construction to operation.

Capabilities need to be verified in practice. I'd like to share two benchmark cases with you.

The first is the Beijing Public Computing Power Platform. This is an AI infrastructure with leading technology, perfect functions, and first - class services built by us for Jingneng Group, providing strong computing power support for the development of the capital's digital economy. In this project, Yingbo Digital undertook the full - cycle services from the early - stage demand analysis and resource planning, to the equipment supply, networking implementation, deployment of computing power and management platforms during the implementation period, and then to the efficient and stable operation services and continuous system optimization during the operation period. This project fully verifies our ability of one - stop full - stack delivery.

The second is the computing power platform for integrated training and inference of a leading large model. This is a landmark project of Yingbo Digital in the field of ultra - large - scale intelligent computing centers. The core requirement of the customer is integrated training and inference - training resources need to be reserved and guaranteed, and inference requires second - level elastic expansion. We delivered a customized computing power cluster with thousands of high - speed interconnected cards, compatible with the customer's self - developed training framework, and provided 24/7 full - link operation and maintenance support.

This project has several key highlights: deployment capabilities covering multiple regions, full - link technical support from deployment to optimization, efficient scheduling of integrated training and inference, 24/7 professional operation and maintenance, and a flexible billing model to help customers significantly reduce computing power usage costs.

These two cases represent our highest level in the two directions of "public computing power services" and "enterprise - level large - model infrastructure" respectively.

Technological innovation cannot be separated from the support of the ecosystem. Yingbo Digital has made in - depth layouts in both the GPU cloud service scenario and industry - university - research cooperation.

In terms of the GPU cloud service scenario, we cover multiple core scenarios such as large - model training and inference, digital marketing, video generation, academic research, and teaching training. Each scenario is matched with the optimal combination of billing models - Reserved ensures resource certainty, Booking meets planned needs, Spot achieves optimal cost, and On - demand provides maximum flexibility.

In terms of industry - university - research ecosystem cooperation, we have sponsored more than 20 academic conferences and carried out scientific research project cooperation with more than 40 universities. We firmly believe that the ultimate competitiveness of computing power infrastructure comes from continuous technological innovation and talent cultivation.

Finally, I'd like to share two judgments of Yingbo Digital on the future of AI infrastructure.

The first judgment is about the industrial trend. The AI industry is moving from a "model - ability competition" to a "computing - power efficiency revolution". The focus of competition is no longer how many parameters your model has, but whether you can support the continuous circulation of massive tokens with lower cost and more stable performance. This is the underlying logic for Yingbo Digital to stick to the infrastructure track.

The second judgment is about Yingbo Digital's mission. We are not just suppliers of "production equipment", but the designers, builders, and formulators of operation and maintenance standards for the entire token factory. Our goal is to become the leader in the transformation from computing power stacking to value creation.

In the wave of the token economy, let's jointly build a computing power foundation that can bear the exponential token demand and jointly promote China's AI industry to move from a "computing - power giant" to a "computing - power powerhouse"!

That's all my sharing today. Thank you for listening! Yingbo Digital looks forward to exploring the infinite possibilities of AI infrastructure with all of you here. If you are interested in our full - stack intelligent computing solution, please scan the QR code to add our staff for further information, or you can communicate with me directly after the meeting. Thank you!