A 0.01-second power outage causes hundreds of thousands of yuan in losses: The most hidden high-profit track in the AI industry is being frantically scrambled for by Chinese enterprises

Jiang Cong, a project supervisor from a leading domestic solar-storage enterprise who is stationed at the Northern Zone of the 24/7 RTC project in Abu Dhabi, has been dealing with a type of "jumping electricity" for the past six months.

The energy storage system he is responsible for debugging connects photovoltaic power generation that fluctuates with cloud cover on one end, and tens of thousands of GPUs (Graphics Processing Units) in the surrounding AIDC computing power (artificial intelligence data center computing power, commonly known as intelligent computing power, different from traditional IDC general computing power) park on the other end.

The AI (Artificial Intelligence) you use every day is backed by a power system that you do not know working at full capacity. Every time you talk to AI, ask AI to generate an image, or make an AI video, tens of thousands of GPUs may be running at the same time behind the scene.

Jiang Cong uses an analogy that the power consumption pattern of these GPUs is different from that in the past. The data centers in the past consume power stably, while the GPUs used for AI have millisecond-level sharp power jumps, like a sports car that steps from idle speed to full throttle and then releases the pedal within one second.

He said, "The power grid is designed for stable power consumption. After the dispatching instruction is issued, the response from capacity expansion at the power plant to the transmission of current to users is counted in seconds." But there is a difference of hundreds of times in between. The power jump of GPUs is counted in milliseconds. Because of this time difference, the voltage drops suddenly, the power grid cannot keep up with the speed, and a 10-millisecond power outage may damage the large model being trained and cause data loss, resulting in a loss of hundreds of thousands of yuan at a time.

What AIDC energy storage needs to do is to fill this "time gap".

It is not a large battery for "storing power for standby" in the traditional sense, but more like a buffer pad in the circuit, which discharges instantaneously to fill the gap within the millisecond-level window when the GPU power rises sharply, and exits smoothly after the power grid completes the adjustment.

This millisecond-level response capability is becoming a rigid demand in the global market.

In the first 5 months of 2026, the global shipment of energy storage installations supporting artificial intelligence data centers reached 10 GWh (Gigawatt-hour), which has exceeded the total shipment scale of the whole year of 2025.

Between June and July 2026, overseas energy storage orders of Tesla, CATL and other Chinese enterprises have been intensively landed. Among them, Tesla reached a 25 GWh super large order with NatPower in June, and cooperated with Esyasoft to deploy more than 15 GWh energy storage systems; CATL signed 7 GWh sodium-ion energy storage orders within one week. The AIDC energy storage scenario is an important incremental market for these enterprises, and Tesla's market share in the field of data center energy storage deployment in North America has exceeded 40%.

Only for the Abu Dhabi RTC solar-storage project in the Middle East, the total energy storage scale has reached 19 GWh, a considerable part of which is reserved for the newly built AIDC computing power park in the surrounding area.

The real demand behind these orders is how to make a large power grid that responds in seconds fully supply power for the "electricity glutton" that jumps in milliseconds.

Orders Are Pouring In

Meng Fan, a salesperson for the European region of a domestic energy storage enterprise, has traveled back and forth between China and Europe for many times this year, connecting with energy investors and computing power park operators in multiple countries. Among the energy storage intention projects he is negotiating, the proportion of clients who actively ask about millisecond-level response, GPU load adaptation and voltage stabilization of computing power machine rooms accounts for most of all the projects he is discussing.

Meng Fan said, "In 2024, I connected with more than 30 intention energy storage orders across Europe throughout the year. Most European energy storage buyers only cared about the consumption of wind and photovoltaic power and peak-valley arbitrage. The former is to make unstable wind and photovoltaic power connected to the grid stably, and the latter is to store power at low price and discharge at high price to earn the price difference. European buyers' core focus is exactly these two major direct economic benefits. By 2025, the proportion of clients who actively mention the demand for energy storage supporting computing power has increased by about half. This year, almost every client will ask about the energy storage scheme for computing power support."

In the first half of 2026, the AIDC energy storage track began to see rapid growth. More than ten enterprises including Sungrow Power, Envision and EVE Energy successively announced new orders in the second quarter, covering multiple markets in Europe, the Middle East and Southeast Asia.

Denmark's KK Group (affiliated to A.P. Moller Holding, the parent company of Maersk, formerly the leading wind power electronic control enterprise KK Wind Solutions, renamed in 2025) has long been engaged in wind power development in Northern Europe, and has started to deploy local AI computing power infrastructure in the past two years.

In the 5 GWh framework agreement that Meng Fan participated in the negotiation, the other party actively allocated 2 GWh to be supplied to the AIDC park in a targeted manner.

The power supply indicators provided by KK Group made Meng Fan realize that this is completely different from the traditional energy storage he has sold for several years. Traditional energy storage is like a domestic water storage tank, which stores and releases water slowly, focusing on large capacity and low cost; AIDC energy storage is more like an emergency team in the circuit. As soon as the GPU "asks for power", the power must be delivered within 10 milliseconds, and any delay may lead to accidents. The PCS (Power Conversion System) power response must be controlled within 10 milliseconds, the continuous discharge rate is required to be at least 2C (the cell can complete the charging and discharging of twice the rated capacity within one hour), the short-term peak value needs to reach 4C (release four times the rated capacity of current in a short time), and the annual system availability is required to reach 99.999%. The conventional wind and photovoltaic energy storage industry generally only requires 99.99%.

Meng Fan gave an example that "In 2024, I could travel all over Europe with one set of technical solutions, but now I have to bring three sets of solutions every time I go out. They correspond to three technical routes of lithium iron phosphate, sodium-ion and sodium-lithium hybrid cells respectively, and I will make recommendations after checking the thermal insulation conditions of the client's machine room on site."

The choices of clients in different climate zones are clearly differentiated.

For AIDC energy storage projects planned in high-latitude areas of Northern Europe, most clients tend to choose sodium-ion batteries.

Meng Fan made an analogy that lithium iron phosphate batteries are afraid of cold, and they tend to "malfunction" when charged below zero degrees Celsius. Lithium precipitates on the surface of the negative electrode, which may reduce the service life in mild cases and cause short circuit in severe cases, and the BMS (Battery Management System) will directly cut off power for protection. Sodium-ion batteries are much more durable. They can be charged normally at minus 20 degrees Celsius, and there is no need to build an additional heating system to "take care of" them.

Clients in southern Denmark and northern Germany prefer lithium iron phosphate batteries, which have lower cost at this stage and their cycle life has been verified by a large number of wind and photovoltaic projects. The temperature in Central Europe is moderate, and both technical routes have their own adopters.

The specifications of orders are changing, and the quotation methods are also changing. The demand structure in Europe, the Middle East and Southeast Asia is also differentiating rapidly.

The total energy storage scale of the Abu Dhabi RTC project in the Middle East is 19 GWh, which is implemented by two leading domestic energy storage enterprises in separate parts. The project bidding document clearly requires that the energy storage should adapt to both photovoltaic peak shaving and AIDC computing power backup.

Khalid, head of energy storage procurement of the Abu Dhabi RTC project, explained to the Economic Observer that for high-power industrial, commercial and computing power projects in Abu Dhabi that are not equipped with self-built energy storage, the grid connection queue of the municipal power grid generally takes 3 to 5 years.

Energy storage can reduce the dependence on the capacity expansion of the municipal power grid. The tiered electricity price for industrial and commercial entities in Abu Dhabi and the monthly maximum demand capacity electricity fee are relatively high. Installing energy storage can discharge power during peak hours, reduce the peak power consumption and cut down the capacity electricity fee expenditure. The UAE also requires that the proportion of green power used in newly built large-scale computing power projects shall not be less than 60%. Photovoltaic power cannot provide stable power supply all day long, so energy storage is a rigid carrier to meet the green power compliance requirements.

Khalid introduced that the three barriers of power grid approval cycle, high electricity fee and green power compliance are restricting the development of related projects, so they decided to integrate photovoltaic energy storage and computing power energy storage into the same bidding project.

The RTC project is also forcing the technical threshold of energy storage products to rise rapidly.

The above Abu Dhabi energy storage enterprise has added new acceptance indicators that do not exist in traditional wind and photovoltaic energy storage in the bidding document. The PCS power response shall be within 10 milliseconds, the seamless off-grid switching time shall not exceed 4 milliseconds, the voltage fluctuation tolerance shall be narrowed from plus or minus 5% to plus or minus 2%, the harmonic distortion rate shall be controlled within 3%, and the annual availability rate shall be increased from 99.99% to 99.999%. All suppliers must complete tens of thousands of load impact simulation tests in the local laboratory in Abu Dhabi, and keep complete waveform reports before they are allowed to enter the site for construction.

Meng Fan said that overseas orders were released intensively in the first half of the year, but there is still a gap between the solutions that clients really want and the standard products running on the production line.

At present, he is sorting out the energy storage parameter templates for different temperature zones in European countries to meet the increasingly segmented customized demand for computing power energy storage.

Delivery Under Pressure

The Abu Dhabi RTC energy storage project signed the supply agreement in the first half of 2026. Lu Chuan is responsible for following up the implementation of the 11.275 GWh bid section of the project. After the contract was signed, he was faced with problems including drawings, production scheduling and operation windows in the early morning.

The project party requires the entire energy storage station to reserve 30% of the capacity for the newly built AIDC computing power park in the surrounding area, and the remaining 70% to ensure all-day peak shaving for photovoltaic power. The power consumption characteristics of the two loads are very different. Photovoltaic power is like cruise control, which charges and discharges stably; the computing power machine room is like driving in a busy urban area, stepping on the accelerator and the brake from time to time, and it requires the energy storage to deliver power the moment the accelerator is pressed, so the drawings must be redone.

Lu Chuan introduced that the earliest general drawings were designed according to the standards of wind and photovoltaic energy storage, with cells arranged and managed uniformly. After the new requirements were put forward, they divided the special computing power cabin into separate areas, re-screened the cells for internal resistance, and concentrated the cells with fast response speed in the computing power area. The drawings have been revised for several versions, and some minor adjustments were made later due to the adjustment of local power access specifications in Abu Dhabi.

The revision of the drawings slowed down the production schedule. Customized electronic control modules and partition wiring harnesses require separate production lines for small-batch flexible production, and the production cycle is nearly two weeks longer than that of standard accessories. The whole batch of goods was divided into several sea shipment batches. The general cabins of the first few batches were all shipped on time, and the cabins with customized accessories in the following batches were delayed for 3 to 7 days respectively.

Lu Chuan said, "We did not delay the shipment on a large scale, but we really cannot ensure that every batch of goods is delivered completely on schedule."

After the goods arrived at the port, the on-site debugging took more time than expected. The maximum daytime temperature in summer in Abu Dhabi from June to August every year can reach more than 45 degrees Celsius, so outdoor operations can only be arranged from zero o'clock to six o'clock in the early morning.

Lu Chuan estimated that the on-site debugging of a conventional overseas energy storage project of the same volume usually takes about 40 days. This project was extended to more than 50 days due to time-limited construction (restricted by overseas power grid and site regulations, construction operations can only be carried out in fixed windows, and continuous construction is not allowed all day long) and customized transformation. The night shift labor cost is 60% higher than that of the day shift. At the hardware level, the computing power partition adds liquid cooling pipeline branches and independent signal isolation modules, so the overall cost of the entire bid section has increased on the basis of the initial budget.

Jiang Cong also encountered similar problems.

The two enterprises adopted different solutions. The enterprise where Lu Chuan works tends to "operate separately", separating the hardware physically between the computing power cabin and the photovoltaic cabin, with clear fault boundaries.

The enterprise where Jiang Cong works adopts cluster-level dynamic partition electronic control. He said that this is like "a big family sharing one bank account". The hardware is not separated, and the energy storage capacity is virtually divided by software programs, and the power is allocated to the side with urgent demand first.

Jiang Cong said, "When the photovoltaic output is sufficient, more power is used for photovoltaic power storage. When the GPU in the surrounding AIDC park has a sharp power rise, the control system temporarily transfers part of the redundant capacity of photovoltaic energy storage to make up for the computing power peak."

The most time-consuming part is how to make the two types of fluctuations "coexist peacefully". The photovoltaic power generation in Abu Dhabi during the day fluctuates with clouds, and there are irregular power spikes in GPU training. When two irregular signals enter the energy storage control system at the same time, the voltage will fluctuate up and down like a roller coaster, which easily causes bus voltage oscillation. What Jiang Cong needs to do is to install a "shock absorber" for this system.

He has carried out multiple rounds of on-site joint simulation debugging, and repeatedly modified the filtering threshold of the EMS (Energy Management System). This type of debugging can only be done in the middle of the night, and a single complete test requires continuous operation for several hours. In a high-temperature environment, the maximum discharge rate of the cell will be passively reduced, but AIDC requires peak high-power discharge. Jiang Cong needs to repeatedly find a balance between safe temperature control and power demand.

He said, "The difficulty of hardware installation is controllable. What really consumes the construction period is the joint fine debugging of the superposition of three types of loads including photovoltaic, energy storage and computing power."

Profit Differentiation

Zhao Haining is tracking the details of electricity bills for several consecutive months. This detail records the monthly power consumption data of an intelligent computing park in Inner Mongolia where he works, from the second half of last year to the first half of this year. Based on this, he calculated a set of capital recovery model "if energy storage is installed".

The park relies on local direct supply of wind and photovoltaic power, and the on-grid electricity price is stable at 0.35 yuan per kWh, with the peak-valley price difference of only 0.22 yuan. Zhao Haining has made several sets of energy storage proportion schemes, covering the overall park, separate buildings and distributed layout.

Calculated with one of the schemes covering the whole park, the total initial investment of the park in equipment, construction and operation and maintenance exceeds 30 million yuan. According to the calculation that the park saves about 4.9 million yuan in electricity fees per year on average, the static capital recovery period is more than six years. If the annual equipment operation and maintenance and cell replacement costs are included, the actual capital recovery period will be longer.

Zhao Haining said, "The capital recovery red line generally accepted by park investors is within five years. If the period exceeds this limit, the capital pressure will increase significantly. A large number of long-term wind and photovoltaic agreements are supported by the West Inner Mongolia Power Grid in the local area of Inner Mongolia. The park can directly lock in long-term affordable green power, and rely on power grid dispatching for supplement when there is no wind or photovoltaic power, so there is no need for energy storage to guarantee power supply. The value that energy storage can create is mainly concentrated in peak-valley price difference arbitrage, and there is no multiple benefits brought by high electricity price and capacity electricity fee in overseas markets."

After internal review, the park management decided to postpone the large-scale energy storage procurement scheme of the park that Zhao Haining participated in. Only the tenants in the park who undertake overseas computing power hosting business have actively installed small energy storage cabins. In the same park, domestic computing power tenants and overseas hosting merchants have significantly different willingness to install energy storage.

The root cause lies in the cost structure.

For domestic computing power projects, the depreciation of GPU and server hardware procurement accounts for 70% to 80% of the operating cost, and the electricity fee only accounts for 5% to 10%. The enterprise's budget is prioritized for computing power hardware, and the investment in energy supporting facilities is placed behind.

Jiang Cong gave an example that this is like spending millions of yuan to buy a sports car. The car cost is the major part, and the oil cost