Last year, over a thousand data center projects were filed in China, with only about 10% actually approved, making direct green power connection a significant bonus. A shift is becoming increasingly clear: AIDC (Artificial Intelligence Data Center) competition is transitioning from resource grabbing and construction racing to green and refined operations. Building the project is just the first step; whether it can go far depends on how well green power is utilized and whether the economics work out. Where should AIDC operators start? Is everything fine once green power is connected? Not necessarily. Without coordinated scheduling, mere physical connection may result in higher green power prices. Can adding energy storage save money? Not necessarily either. In actual calculations for some projects, if energy storage is only used for generation-side regulation, the real economic value released may be less than 10%. Power generation, load, electricity price, and energy storage operating in silos is precisely the crux of extensive operations. What needs to be solved in the second half of AIDC is not just whether there is power or energy storage, but whether these resources can be integrated into a single decision-making system to find optimal solutions amid continuous fluctuations. Only by truly playing the chess game of source, grid, load, and storage can the economic viability of AIDC be guaranteed. In this issue, Ai Fenxi interviews Jian Yumin, CEO of Damiao Intelligence, to discuss where the economic benefits of AIDC will ultimately come from as it enters the stage of refined operations, covering dimensions such as AIDC project implementation, computing-power and electricity coordination, energy storage value, and competitive landscape.
Key Takeaways
- Energy is the core bottleneck for future AIDC construction.
A large number of AIDC project applications have been rejected globally, with one of the main reasons being the inability to obtain energy quotas and grid connection approvals. In the wave of building a new power system and power market reform in China, AIDC projects have been deeply integrated with new energy, even prompting energy companies to build their own computing centers to unlock the full-chain revenue "from power to computing power."
- AIDC is shifting from a real estate logic to an operational logic, with computing power consumption and green power absorption capabilities being the lifeline.
Only projects with truly stable customers and pre-consumption contracts can pass approval; the simple "build first, find customers later" model has been systematically suppressed.
- The western computing power landscape has differentiated, with Ulanqab in Inner Mongolia becoming a benchmark, while Xinjiang still has significant room for growth.
Ulanqab's success stems from three reasons: low-latency networks, a 2-hour ecosystem circle from Beijing, and real computing power consumption.
- The essence of electricity price prediction is market game decision-making.
As fully market-based power trading opens up, electricity prices depend not only on physical factors like weather but also on the emotional gaming of traders. One of the current technical routes is shifting from pursuing relatively high prediction accuracy to quantitative trading strategies and game algorithms based on model fusion optimization.
- Energy storage is not just a supporting facility; it can also become a regulating resource to improve system operation efficiency.
According to Damiao Technology's calculations for some western projects, under the premise of complying with local power market rules and access conditions, using energy storage to smooth new energy output, optimize power trading, and optimize user-side basic electricity fees is expected to increase the overall project IRR (Internal Rate of Return) by 1 to 1.5 percentage points.
- The AIDC land grab will end in a year, and competition will shift from land grabbing to operations.
The current stage is about energy providers locking in customers; the next stage will enter system competition centered on scheduling capability, trading capability, data capability, and algorithm capability.
The following is the transcript of this interview, slightly modified without changing the original meaning.
01. AIDC Project Window Guidance is Hard to Pass, There Are 3 Core Bottlenecks
Ai Fenxi: Recently, major energy companies have been actively deploying AIDC. Will the core barrier of the future AIDC industry chain increasingly tend towards energy?
Jian Yumin: In the future, the biggest bottleneck in the entire AIDC and even the AI (Artificial Intelligence) industry chain will be energy. AIDC construction might even fail to get approval without obtaining new energy quotas. We can look at a set of data: last year, the approval rate for computing center applications across China was about 10%. In the US, this year has also started requiring computing centers to self-configure new energy. Globally, rejected computing centers account for a large proportion, and the primary reason is the difficulty in resolving energy supply issues. The core reason energy companies deploy computing centers is their long-term, stable power demand. As power marketization advances, energy companies not only need to generate power but also find continuous and stable consumption scenarios, and computing centers are exactly one of them. Meanwhile, models like direct green power connection are integrating new energy supply and computing power loads more tightly. If energy companies form long-term cooperation with computing centers and even deploy computing services, they will have the opportunity to connect power generation, consumption, and computing operations, expanding full-chain value under compliance premises, rather than relying solely on power sales for revenue.
Ai Fenxi: Besides energy, what are the other core bottlenecks?
Jian Yumin: There are two other core reasons projects get blocked: first, whether the applicant truly possesses computing power consumption capability; second, whether it can align with the domestic substitution strategy. Currently, many computing power companies apply for construction because they see market space and want to build houses in advance. China's computing centers already experienced such a "real estate" wave in the previous stage, resulting in computing centers failing to generate very good profitability after construction. During this window guidance approval process, the government on the one hand has put forward high requirements for the professionalism of the operating entity, explicitly requiring a professional operator as the lead entity. On the other hand, it requires corresponding customer pre-consumption contracts throughout the application process. For example, a large number of projects were applied for in a western province last year, but ultimately failed to get approved. The biggest problem lies in the end-terminal consumption users. The second bottleneck is domestic adaptation. For projects involving central and state-owned enterprises, domestic substitution is an important requirement.
Ai Fenxi: Is it harder for retail computing centers to get approved?
Jian Yumin: Applications for retail customers are indeed relatively difficult, mainly depending on customer reservations. In contrast, operators find it relatively easier to apply for projects because they inherently undertake the task of absorbing the national computing power development, so they basically have some reserved customers to make commitments. Some computing centers providing commercial services for internet giants can also basically get approved when applying for projects.
Ai Fenxi: Why has there been a clear differentiation among western computing power nodes?
Jian Yumin: Whether western nodes can form scale depends corely on the network, service ecosystem, and local computing power consumption capability. Among the current several western nodes, Ulanqab has developed relatively faster, mainly for three reasons.
First, the network infrastructure is relatively mature. Ulanqab completed network node construction early in the promotion stage of the "East Data, West Computing" project. Its low-latency network makes it easier to undertake latency-sensitive computing power businesses. The rapid customer adoption after local cabinet deployment also shows that for western nodes to form scale, network conditions and real consumption demand must be in place simultaneously.
The second reason is the 2-hour ecosystem circle. Users have a basic requirement for all computing centers: after an incident occurs, personnel need to arrive within two hours, and fault recovery needs to be achieved within four to eight hours. Ulanqab is within the 2-hour ecosystem circle of Beijing, so there will be no deficiency in its ecological positioning. Meanwhile, with the initial ecological coverage, a complete service ecosystem chain has been established in this area subsequently.
Third, local businesses can provide startup support. For western nodes to form scale, the key is to first introduce sufficient computing power demand. Taking Gansu as an example, during the initial stage of the project, the government and operators jointly introduced large customers to form the first batch of computing power consumption, after which the local ecosystem development was gradually driven.
02. Incremental Profits of AIDC are Hidden in Electricity, the Key is Game Algorithms and Energy Storage Facilities
Ai Fenxi: What predictions need to be made for computing-power and electricity coordination? How much gap can be created among different vendors?
Jian Yumin: We cannot talk about predictions in general; it must be divided into at least three categories: power generation prediction, load prediction, and electricity price prediction. Power generation and load predictions mainly face the uncertainty of the physical world. Whoever has more complete data and a more reasonable data structure has a better chance of achieving higher accuracy. Taking load prediction as an example, our accuracy in the 24-hour dimension can reach over 92%, and ultra-short-term can reach over 98%. Different projects are affected by data completeness and operating condition complexity, leading to gaps in results. As the complexity of inference tasks increases and agent tasks multiply, AIDC loads become harder to predict. Complex inference splits a task into sub-tasks with different time sequences, and agents have stronger personalized characteristics, so load randomness will continue to rise. The logic of electricity price prediction is different; besides being affected by supply, demand, and weather, it also includes the trading behavior of market entities, so it cannot be judged solely by the thinking of the first two types of predictions. In terms of difficulty, the decision-making in power trading is the most critical and also the most difficult.
Ai Fenxi: Why is power trading decision-making the most difficult and critical?
Jian Yumin: Because what truly determines revenue is not just prediction, but how to make decisions after prediction. Electricity price is not a simple natural variable; it is also affected by the quotation strategies and trading behaviors of market entities, so it is unrealistic to want to achieve 100% prediction accuracy. As the degree of power marketization deepens, quotations from the generation side and user side may have temporal and spatial mismatches, and electricity price fluctuations will become more complex. This is actually a multi-party game in the market. Therefore, the higher the degree of marketization, the harder it is to predict electricity prices. For computing center operators, what matters is not pursuing an absolutely accurate number, but formulating scheduling and trading strategies that better balance revenue and risk through decision intelligence based on uncertain prediction results.
Ai Fenxi: What specifically does this game refer to?
Jian Yumin: The game here mainly refers to the emotional game among multiple entities involved in power trading markets, including generation, sales, and consumption. In electricity price prediction, even leading teams and companies in the industry have relatively large deviations in prediction accuracy, with the lowest periods even dropping to 40%. What is the reason? The formation mechanism of electricity prices itself contains the emotions of the entire power market traders. Before full marketization, the formation mechanism of electricity prices was mainly a planned track mechanism led by the power grid, making electricity prices relatively easy to predict. However, after entering full marketization, you will see the fluctuation patterns of electricity prices becoming increasingly strange, which actually reflects the game trends among various trading entities. The higher the degree of marketization in the future, the less accurate electricity price predictions will be, so the technical route we take is to balance both prediction and decision-making routes. At the current stage, we will fuse multiple different electricity price prediction models. Through weight optimization, we can find relatively more accurate weighted values in different regions, time windows, and weather environments to cope with the current transition stage. In the next stage, when the entire full-scale marketization is determined by human factors, users' trading emotions can reflect more value. This part of the game is to find a better and more precise trading trend within the trading value distribution of each trading entity, which is also the second step of optimization we need to do, considering how to trade in the market space based on quantitative logic.
Ai Fenxi: Against the backdrop of inaccurate electricity price predictions, is Damiao purely playing for the win rate like quantitative investment, or playing for the odds?
Jian Yumin: There is no consensus yet on which game method is currently used in China's power market. There is an example overseas, a company called Octopus Energy, which plays for the odds. Specifically, what Octopus Energy plays for is the time period when the entire electricity price market prediction is the least accurate, and then when everyone is inaccurate, they do reverse arbitrage. In this case, they can earn relatively more money. Actually, there is a simple case in China: an old lady sweeping the street at the entrance of the stock exchange hall. When everyone goes in to buy, she doesn't go in; then when no one goes in, she goes in to buy. It's this kind of thinking logic.
Ai Fenxi: On the electricity price prediction end, subsidiaries of the power grid possess more data. Is it easier for them to build competitive barriers?
Jian Yumin: This advantage will definitely exist, but we can integrate their advantages through a model fusion mechanism. If the power grid's electricity price prediction capability can be improved in the future, the power grid's data itself will also be a key focus of our fusion model. In fact, currently, power grids in all different provinces have their own prediction data published online. The value of model fusion is to use multi-model fusion to reduce the disturbance of a single model. The current domestic power market is still steadily improving, and the performance of power grid public data and different prediction models will vary across different regions, time periods, and weather conditions. The system will dynamically compare each model, increase the weight for models that are more adaptable to the current scenario, and then fuse them with other models to obtain more robust results.
Ai Fenxi: At the policy level, will this kind of electricity price trading strategy be restricted? For example, to prevent excessive speculation and electricity hoarding behaviors.
Jian Yumin: Policies will definitely have restrictions. For example, corresponding to buying and selling in the power market, there is an upper limit suppression in China. If electricity sales companies are engaged in electrical energy trading itself and their profits exceed a certain percentage, which varies by province, for instance, in Inner Mongolia, if profits exceed 10%, there will be an excess recovery mechanism that returns the earned money to the power generation side. This is also to prevent behaviors like electricity hoarding affecting actual market power supply, as seen in the US.
Ai Fenxi: Currently, the power market has a 15-minute settlement cycle. Will this granularity continue to drop to 5 minutes in the future? Can AI assist, and what is the limit after assistance?
Jian Yumin: From my personal cognition, 5 minutes is the limit. 15 minutes is the international unified standard because the concept of energy block trading is based on the foundational theory established at MIT in 1881, and the world follows the 15-minute dimension. To achieve real-time balance through trading, it is possible to reduce the granularity further. For example, the trend of 5-minute settlement has already appeared in Shanxi's settlement. However, whether it will be promoted nationwide, we think not necessarily. Because the finer the granularity, the higher the level of complexity improvement for trading, it is impossible to achieve this level solely by human power; AI can play a role. But AI assistance also has its limits, affected by multiple factors such as data collection, inference, and execution. At the current stage, 5 minutes is basically the limit for underlying settlement and AI response.
Ai Fenxi: Predictions for power generation and load are based on objective facts and industry experience. If agent technology fully enters the prediction end and self-evolves, can latecomers break the industry barriers of first-movers?
Jian Yumin: This depends on a key assumption: whether anyone will actually do computing-power and electricity coordination to structure industry experience and use it to train general large models. The self-evolution capability of agents is actually based on the capability of the large language model itself. For example, now we use agents to do calculation models, which can be much better than traditional solutions. But the core point in this process lies in the modeling capability of the large model itself; it needs to have foundational knowledge in mathematics, physics, and energy. This is the most basic logic of the energy full-chain model we are currently building. The fine-tuned energy large model, in the vertical scenario of computing-power and electricity coordination, is much better than general large models in both modeling and computing capabilities. If a general large model wants to build such a capability, especially in the field of computing-power and electricity coordination, it first needs to achieve the same capabilities as vertical models on the general large model and create its profit model. I think this is a relatively big challenge. Another point is that when the self-evolution capability of agents reaches a certain level, the cost is disguised and transferred, converting our endogenous industry experience capability into exogenous actual token consumption. In this case, whether agent optimization based on general large models can reduce overall operating costs more than vertical large models needs to be demonstrated.
Ai Fenxi: How do you view the value of energy storage facilities in computing-power and electricity coordination?
Jian Yumin: Energy storage facilities are the largest value assets in source-grid-load-storage and direct green power connection. In the cognition of traditional energy enterprises, the role of energy storage is singular, which is to improve new energy absorption capacity, such as extending the hours of wind and solar power generation or smoothing the output curve. But if only this step is taken, the utilization rate of energy storage is actually very low. We have calculated in some projects that if configured according to standard settings, such as a 4-hour, 15% load energy storage system, and it is only used for generation-side regulation, the real economic value it exerts may be less than 10%. That is to say, energy storage is idle 90% of the time. But the problem is that energy storage itself is a two-way asset; it is not just limited to "charging in and discharging out." Essentially, it is a dynamic resource that can participate in power system operation and regulation.
Ai Fenxi: That is to say, energy storage actually has ignored usage methods?
Jian Yumin: Many people only see the generation-side value of energy storage but ignore its value on the grid side and user side. If energy storage is placed in the entire system, it has three types of value. First, smoothing new energy output; second, optimizing energy costs through peak-valley price differences and power market trading; third, optimizing user-side basic electricity fees. For computing centers implementing a two-part tariff, basic electricity fees can be charged based on transformer capacity or maximum demand. By using energy storage to peak-shave and valley-fill, and controlling maximum demand, this part of the cost can be reduced under the premise of ensuring safe operation. We calculated in some western projects that, under specific configurations and operation simulations, after revitalizing the idle capacity of energy storage, the overall project IRR is expected to increase by 1 to 1.5 percentage points.
Ai Fenxi: Energy storage seems to be no longer just a supporting facility, but a core variable that can participate in scheduling and trading.
Jian Yumin: Energy storage does not exist independently; it must be optimized within the entire system. Its scheduling objects include three dimensions: the uncertainty on the generation side, the fluctuation of market electricity prices, and the real-time changes on the load side. Only by putting these three into a single framework for global optimization can the value of energy storage be truly released.
Ai Fenxi: Besides new AIDC projects, can computing-power and electricity coordination get a share of the pie from the upgrading and renovation of old computing centers? Or are there hard restrictions?
Jian Yumin: There are no hard restrictions. As long as old computing centers can find a way to invest and build some regulating resources, they can also participate in the computing-power and electricity coordination market to share profits. As long as built computing centers are configured with energy storage to give them regulating capabilities, they can participate. For example, in the Lingang project, it actually configured energy storage in an already built computing center. Now policies are also exploring whether existing large loads can sign long-term agreements with some new energy project entities, or whether they can disconnect and reconnect to the grid. These all have certain room for development. However, there is indeed a problem regarding the capability of investment and construction and the issue of disconnecting and reconnecting to the grid. In the eastern regions where land is extremely expensive, due to site limitations, the difficulty of investing and building energy facilities is indeed relatively large, which is an undeniable fact.
Ai Fenxi: Since computing-power and electricity coordination is a systematic project spanning multiple ends of source, grid, load, and storage, does it mean that it cannot be established as an independent project, but must be directly included in the EPC (Engineering, Procurement, and Construction) general contracting project of green power or source-grid-load-storage integration?
Jian Yumin: The industry volume of computing-power and electricity coordination is very large, and it is indeed not established as a computing-power and electricity coordination project, because the declaration caliber of our project establishment is currently still computing infrastructure construction. It is divided into two types: one is the project establishment of source-grid-load-storage integration, and the other is establishing the computing center and energy construction on two separate sides.
03. AIDC Land Grab is About to End, Competition Focus Shifts to Computing-Power and Electricity Coordination
Ai Fenxi: What is the development trend of the computing-power and electricity coordination market?
Jian Yumin: We estimate that the computing-power and electricity coordination market will enter a more intense competition stage in the second half of 2027. Currently, energy providers are grabbing land to lock in computing power consumption customers. The competition focus of energy providers is not on computing-power and electricity coordination, but on green power absorption. By the second half of next year, when these computing power customers have established binding relationships with energy providers, the remaining competition will concentrate on service providers like Damiao. During this period, major energy companies will also consistently accumulate operation data to improve their scheduling capabilities in computing-power and electricity coordination. This data accumulation is very important because for energy enterprises to do a good job in computing-power and electricity coordination, they still need to make up for the energy consumption data, load characteristics, and user-side operation experience of computing centers. In the past, generation-side enterprises mainly sold power to the grid and still had a certain distance from end-users. Such capabilities need to be gradually accumulated through real projects and long-term operation data. Computing centers are familiar with business loads, and energy enterprises are familiar with the generation side. If the data and experience on both sides are not connected, it will be difficult to quickly improve prediction and scheduling effects. For professional service providers, the real opportunity lies in incorporating the data of source, grid, load, and storage into a unified management system and transforming predictions into continuous scheduling and operation capabilities. Of course, Damiao Technology is also increasing R&D efforts this year while continuously establishing deep connections with customers.
Ai Fenxi: Are internet giants the main competitors? For example, choosing to develop systems in-house and no longer needing third-party service providers? How should third-party service providers respond?
Jian Yumin: The core point of third-party service providers' development lies in steadily advancing their own professionalism. In-house development by giants often faces the problems of overfitting or mismatched business focus. For example, all computing centers will use a set of systems—the Power and Environment Monitoring System. The technical threshold is not high, mainly involving data collection and fault alarm at the IoT (Internet of Things) layer. Currently, the power and environment monitoring market has formed a group of highly specialized suppliers. Giants can build very strong systems, but they cannot necessarily turn them into industry-general products. A certain internet giant once developed a power and environment monitoring system in-house. The system's penetration capability is very strong, and it is also one of the most excellent platforms we have seen in the industry. However, such systems are usually polished around their own machine room environments and management specifications. When switched to other scenarios, the adaptation and customized development costs will significantly increase. Coupled with the competitive relationship among cloud business vendors, their internal systems are also difficult to directly promote to peers. Therefore, the value of third-party service providers still lies in general product capabilities and cross-scenario experience. In addition, the core business of giants is not on the physical assets of computing centers; computing centers are just the foundation for giants. For example, although a top internet enterprise's current attitude towards computing centers is self-construction, the target of bidding is actually IDC (Internet Data Center) Colo (Colocation) providers. If the general contracting bidding targets for computing centers are already at such a professional level, then the underlying operation and maintenance systems are even less likely to be developed in-house, which does not conform to the logic of business focus.
Ai Fenxi: Besides internet giants and energy behemoths, what other segmented tracks will computing-power and electricity coordination evolve into?
Jian Yumin: The boundary of computing-power and electricity coordination is very broad, and its core goal is actually just one thing: how to reduce token costs. Deduced from this point, the entire industry chain has been segmented into many tracks. On the infrastructure side, for example, in scenarios like direct green power connection and DC microgrids, DC-side equipment has become a focus of industry development direction, which has already formed a new track. Moving up is the prefabricated complete supporting equipment supplier track. The core of computing centers is to serve chips, and chips have their own life cycles. After each generation of chips is researched and produced, the earlier they can be put into use in computing centers, the more value the chips can exert in their life cycles. Therefore, prefabrication will serve as a way to quickly invest and build computing centers, also generating a new track. Recently, various giants are doing prefabricated in-house development, especially for overseas computing centers, where the proportion of prefabricated in-house development will be very high. Besides prefabrication and chip competition, it also includes ecological software layer competition like inference acceleration and computing power scheduling. Further up is the optimized utilization of the entire token. In the agent era, all functions are handed over to large models, and the token consumption is very large. So now many AI infra (Infrastructure) enterprises are also exploring and optimizing the "computing power-task" allocation. For example, when complex problems occur, when to use local small models and when to throughput these tasks to large models for cloud token consumption.
Ai Fenxi: Damiao Intelligence released four agents in the first half of this year, covering electricity price prediction, load prediction, scheduling regulation, and trading settlement. What is the core difference between these agents in actual implementation compared to the past?
Jian Yumin: The core difference lies in the fact that we can truly make it save money in direct green power connection scenarios, rather than costing money. In traditional direct green power connection or source-grid-load-storage projects, because self-generated new energy highly overlaps with the low-price time of grid power, rigid physical connections will conversely raise the user's final settlement electricity price. Combining actual cases, we can see that if simple direct interconnection is adopted without deep scheduling control, the electricity price may be conversely pushed up to 1.3 to 1.5 times. In western provinces, the settlement electricity price is even higher than in Shanghai. This is the reason why many projects now look good but are not practical. With our multiple agents conducting overall energy scheduling and trading optimization, finding the optimal solution through algorithms amid complex electricity price and load fluctuations, this is the core value of our agents. At the World Artificial Intelligence Conference held in July this year, we also officially released the Computing-Power and Electricity Coordination 2.0 platform, which includes the energy large model and agent matrix. It can achieve multi-source spatiotemporal and multi-market data prediction and trend perception, as well as global coordination and multi-objective optimal strategy solving. Different from the single-point capability verification of 1.0, 2.0 has completed the upgrade from the integrated closed loop of "prediction-scheduling-trading" to the intelligent operation of the full life cycle of AIDC direct green power connection, integrating scattered agent capabilities into an integrated operating system. We also recently shared our industry understanding of how direct green power connection can go from "losing money" to "making a profit." Its essence is the lack of multi-dimensional collaborative optimization capabilities for computing power loads, new energy output, and spot electricity prices. Computing-Power and Electricity Coordination 2.0, through multi-source prediction fusion and joint optimization of computing power, electricity, storage, and carbon, transforms direct green power connection from a compliance investment into a source of operational revenue, which also represents the industry's deep transformation from "focusing on hardware construction" to "focusing on operational benefits."