The rapid expansion of the artificial intelligence industry has led to a continuous increase in the electricity demand of data centers, resulting in an explosive growth in power consumption. At the same time, the high-speed operation of AI training and inference tasks has significantly intensified the fluctuations in data center loads. Compared with traditional industrial loads, AI computing clusters have higher requirements for power supply stability, continuity, and response speed. All these are driving the acceleration of the iterative upgrading of power infrastructure.
The state of Texas in the United States (hereinafter referred to as “Texas”) is a typical example of this round of industry transformation. The local penetration rate of new energy ranks among the top in the United States, and it also has a large number of newly built AI data centers. The fluctuating output of new energy combined with the highly volatile AI power load has made the Texas power grid no longer a simple transmission channel; instead, it is gradually becoming a practical testing ground for adapting to high computing power loads in terms of power stability.
In the past two years, there has been a boom in the construction of backhaul power (BTM) facilities in Texas. According to statistics from the research institution Cleanview, from 2024 to 2025, the officially announced backhaul power data center projects in the United States were approximately 56 gigawatts in scale, with over 20 gigawatts of them being located in Texas; in the first four months of 2026 alone, Texas added another 10 gigawatts of related projects.
At present, the construction pace of AI data centers is significantly faster than the expansion speed of traditional power grids.
Although the Texas grid operator ERCOT is speeding up the approval process for large-scale power load integration, the new transmission lines will not be able to be put into operation gradually until early 2030 at the earliest. This means that projects that rely solely on the public grid for integration will have their construction schedules forced to be postponed by several years.
Due to the mismatch between supply and demand, a completely new power supply model is gradually taking shape. Data center parks build their own natural gas power stations and equip them with energy storage devices, achieving partial off-grid operation. Subsequently, they will be connected to the public main power grid. This approach is called “bridge-to-grid” in the industry. During this industry transformation, the industry positioning of energy storage has also undergone a fundamental change.
In the past, energy storage was often regarded as a supporting component for new energy power generation, mainly undertaking auxiliary tasks such as frequency regulation and peak shaving. Nowadays, energy storage has gradually become an indispensable basic configuration in the AI power system. Whether it is gas power generation combined with energy storage, long-duration energy storage for wind and solar power stations, or the installation of battery systems in grid-connected data centers, energy storage has transformed from an optional component to a core device ensuring the stable operation of computing power.
High-density computing power forces the power grid to raise its entry standards.
The AI equipment has extremely strict requirements for power supply stability, which is the core reason why energy storage has become a standard feature for local data centers. The training cycle of an AI large model can last for several weeks or even months. Even a millisecond fluctuation in voltage or frequency can cause training to be interrupted, server tripping, and all the previous training data may become invalid, requiring a reiteration of the training process. The grid fluctuations that ordinary commercial and industrial facilities can tolerate would be considered a major operational accident in a computing power center.
Due to the unique characteristics of computing power load, ERCOT has introduced stricter grid connection rules for high-power electricity consumption loads. In 2025, Texas is promoting the SB6 bill around the control of large new loads, authorizing ERCOT to strengthen the grid connection management of large new loads, and requiring facilities with a capacity of 75 megawatts or more to have stronger grid disturbance tolerance. Facilities that do not meet the requirements may be prioritized for load shedding in case of grid emergency.
After the policy was implemented, the cost of grid connection for AI data centers increased, and they also had to install power support equipment for voltage stabilization and frequency stabilization. For GPU clusters with intense load fluctuations and a scale of several hundred megawatts, relying solely on the public grid can no longer meet the power supply requirements. Coupled with the actual problems such as prolonged grid connection queues and delayed grid expansion, more and more data center developers choose to build their own power supplies and adopt off-grid or semi-off-grid power supply solutions.
The AI data center itself is a typical unstable power load. When the GPU cluster is working at high intensity, the fluctuation of power consumption is much faster than that of traditional industrial equipment. A data center with a capacity of hundreds of megawatts can experience significant power fluctuations in a short period of time. The most direct solution to this problem is to configure energy storage. Through millisecond-level response to stabilize power changes and maintain voltage and frequency stability, it can also provide short-term backup power supply in extreme conditions. Among them, the technical characteristics of grid-connected energy storage are highly compatible with the power supply requirements of the AI industry.
Traditional grid-connected energy storage relies on the existing power grid for operation, while grid-forming energy storage can autonomously establish voltage and frequency references even in weak grids or even in a partially disconnected state, and has stronger system support capabilities. It is widely believed in the industry that AI data centers with significant load fluctuations will become the main application scenario for grid-forming energy storage.
Three mainstream power supply modes: Integrating energy storage deeply into the computing power system
At present, the AI data centers in Texas mainly adopt three types of power supply solutions. Although the construction paths are different, the common trend is already quite obvious. Energy storage is no longer just an auxiliary device in the new energy system, but is becoming a core component of the computing infrastructure.
Among them, the “bridge power supply” model combining natural gas and energy storage is the fastest to be implemented and has the highest adoption rate in Texas at present. The reason why this model was the first to be implemented in Texas is that Texas has energy conditions that other major data center clusters in the United States do not possess.
Compared with the traditional data center clusters in Virginia and Pennsylvania under the jurisdiction of PJM, Texas has more abundant and lower-cost natural gas resources. Especially, West Texas is adjacent to the largest shale oil and gas production area in the United States, the Permian Basin. There is a continuous supply of associated gas, and the gas price is significantly lower than that in the northeastern part of the United States. At the same time, the approval process for natural gas power generation in Texas is relatively lenient, with low land prices and easy access to industrial land, enabling the rapid deployment of large-scale gas turbines. Although the PJM region has relatively more abundant grid capacity, it is affected by factors such as environmental protection approval and cross-state pipeline restrictions, resulting in longer and higher-cost cycles for new large-scale gas power projects.
In contrast, although California has the world’s largest technology cluster, it has long faced issues such as the retirement of natural gas power plants, stricter environmental policies, and a tight power market capacity. It is difficult for AI data centers to build large-scale new gas-powered backup power sources. The local area relies more on expensive grid electricity and long-term power purchase agreements (PPA), lacking the practical conditions like those in Texas where power sources are built first and then connected to the grid.
This makes the state one of the few regions in the United States that can simultaneously meet the three conditions of low electricity prices, high flexibility, and rapid deployment. For AI data centers, this is particularly crucial. Because for AI loads, what is valued is not an absolutely low electricity price, but stable, continuous, and controllable power supply capabilities.
The natural gas generator set precisely possesses this characteristic. Compared to wind power and photovoltaic energy, it offers greater flexibility in scheduling, stronger stability, and a smaller footprint, making it more suitable for the continuous operation of high-density AI data centers. Meanwhile, the energy storage system assumes another crucial function. It acts as a buffer between the natural gas generator set and the AI load, used to balance load fluctuations, stabilize voltage and frequency, and undertake tasks such as black start, instantaneous backup, and power switching. As the scale of the GPU cluster continues to expand, the energy storage has actually begun to assume some of the functions of traditional UPS (uninterruptible power supply) systems and backup power sources.
The Matador project under construction by Fermi America, a private energy developer in the United States, is a typical example. The total planned scale of this energy park reaches 17 gigawatts, of which 11 gigawatts are for natural gas power generation, along with 4.4 gigawatts of nuclear power, photovoltaic and energy storage equipment. Essentially, it is already close to a large self-owned power plant. The project has obtained the first approximately 6 gigawatts of natural gas power generation emission permits and has initiated the purchase of gas turbines and the preliminary design of nuclear power, but it is still in the early development and financing stage and has not yet been officially put into operation.
Another representative case is the AI data center project jointly promoted by the artificial intelligence data center developer Prometheus Hyperscale, the Texas independent power producer Conduit Power, and the French utility company ENGIE. This project adopts a bridging power supply mode of “natural gas + energy storage”. In the early stage, the self-built power supply directly powers the data center. Once the transmission facilities of ERCOT are completed, it will be connected to the public power grid. After being connected to the grid, this system will not be withdrawn but will continue to exist as a backup power supply and flexible resource. It can even sell electricity back to ERCOT in reverse. A single park can provide approximately 300 megawatts of transitional power supply capacity.
The core of this model is not merely building its own power supply, but rather transforming the traditional approach of waiting for grid connection into first laying out computing power and then waiting for the grid. When the construction period of data centers is faster than the expansion speed of the power transmission system, Texas, with its low-cost natural gas, mature oil and gas infrastructure, and relatively lenient energy regulatory system, is becoming the earliest large-scale experimental field for this new power supply model.
The second mainstream solution is “grid connection + energy storage”. The data center retains the public grid connection method while also equipping large-scale energy storage devices. The core idea is not to avoid the public grid, but to optimize the power consumption characteristics through energy storage, so that the highly fluctuating AI load can adapt to the operation rhythm of the grid.
This model rapidly took root in Texas, closely related to the operational characteristics of the ERCOT market. Over the past decade, the installed capacity of wind power and photovoltaic power in Texas has grown rapidly, and the proportion of renewable energy has continued to increase, gradually shaping ERCOT into a typical high-volatility power market.
However, ERCOT does not have a mature capacity market like PJM. It still operates as a pure energy market (energy-only market), and the profitability of energy storage projects heavily depends on real-time price fluctuations and changes in the prices of ancillary services. The large-scale integration of AI data centers has further exacerbated this volatility.
In this situation, energy storage is evolving simultaneously at both the technical and financial levels. At the technical level, it responds quickly to load changes, participates in frequency regulation, peak shaving and valley filling, voltage support, and reserve capacity adjustment, buffering the impact of AI loads on the power grid. At the financial level, energy storage gradually evolves into a tool for managing electricity prices, helping data centers lock in electricity costs and hedge against the sharp real-time price fluctuations in the ERCOT market.
The Hidden Lakes energy storage project, jointly developed by the US energy storage developer GridStor and the European energy trader Axpo, is a highly representative case of this model. The project is located in Texas, with a planned capacity of 220 megawatts/440 megawatt-hours. It reached a long-term revenue swap agreement in 2025 and is scheduled to be operational in 2026. The project is managed by GridStor for the construction and holding of the energy storage assets, while Axpo is responsible for the subsequent market transactions, revenue management, and risk hedging of the project. Compared to traditional long-term power purchase agreements, this revenue swap model is closer to a financialized energy storage structure.
Since the energy storage revenue in the ERCOT market is completely exposed to market fluctuations, the project’s profitability usually relies on the arbitrage logic of charging at low prices and discharging at high prices, as well as participating in the ancillary service market to obtain additional income. Therefore, the cash flow is highly uncertain. The revenue swap mechanism essentially redivides this risk. GridStor hopes to obtain long-term, stable, and predictable infrastructure revenue, while Axpo is willing to bear the price fluctuation risk in the ERCOT market. Specifically, Axpo will provide GridStor with pre-agreed stable income through financial contracts, and the actual trading income of the energy storage project in the ERCOT market will be obtained and managed by Axpo. If the market fluctuations are severe enough, Axpo can obtain excess profits; otherwise, it needs to bear the market risk on its own.
This also enables Axpo to play a role in the entire project that goes beyond the traditional definition of a power trader, and is more akin to a market maker and risk underwriter for energy storage projects. On one hand, it utilizes its long-established trading capabilities in electricity, commodities, and natural gas in Europe and North America to conduct real-time arbitrage and risk management in the ERCOT market. On the other hand, it is actually providing a function similar to “yield insurance” for energy storage projects, helping infrastructure investors reduce the uncertainty of future cash flows.
For AI data centers, the significance of this model is rapidly increasing. Because what these computing parks really struggle to cope with is not the relatively high average electricity price, but rather the sharp fluctuations in electricity prices and the uncertainty of power supply. Therefore, energy storage is no longer just a simple hardware device; it has begun to transform into a form of stability service. GridStor has also clearly stated that the future service targets of the Hidden Lakes project include not only utility customers but also large data center loads. This means that the financialization of energy storage is beginning to deeply integrate with the electricity consumption demands of AI, and data centers are no longer purchasing only batteries themselves, but a complete new energy service system built around stable power supply, real-time regulation, and price risk management.
The third mode is the green electricity computing power solution combining new energy with long-term energy storage. Leveraging the low-cost renewable energy resources in Texas, combined with cross-day-level long-term energy storage, a relatively independent and stable power supply system based on new energy is established to provide clean electricity for data centers.
In this model, wind power and photovoltaic energy play the role of basic power generation. However, due to their strong natural volatility, they must rely on ultra-long-term energy storage systems to smooth out energy across different time periods, enabling power supply to extend from hourly balance to daily or even multi-day balance. Currently, this model is still in its early stage of development, but representative projects have already emerged.
In 2024, the US long-duration energy storage company Form Energy entered into a partnership with the AI infrastructure company Crusoe, planning to provide a battery energy storage system of up to 12 gigawatt-hours for Crusoe’s data center project. This partnership is regarded as the first time that long-duration energy storage has entered the commercialization and implementation stage in the form of an exclusive energy system for data centers.
The biggest difference between this project and traditional energy storage projects lies in the fact that it does not provide auxiliary services for the power grid, but directly serves a single computing load. Crusoe locks in the storage capacity in advance and integrates it as part of its own data center energy system, thereby physically securing the long-term power supply capacity in the future and transforming the energy storage from a power market asset into a part of the computing infrastructure.
In terms of the technical approach, the key to this model lies in the introduction of long-duration energy storage technologies such as iron-air batteries. Currently, the mainstream lithium-ion battery energy storage systems typically only have a discharge capacity of 2 to 4 hours, and are mainly used for frequency regulation and short-term peak shaving. In contrast, iron-air batteries can achieve a continuous discharge capacity of up to approximately 100 hours, possessing cross-day-level energy scheduling capabilities. This capability enables it to continuously provide stable power to data centers when wind and solar power output is insufficient or during extreme weather conditions, thereby physically breaking the intermittent constraints of renewable energy generation. This energy-storage-power-computing power supply system also provides a feasible path for future fully green electricity-driven computing infrastructure.
Evolution of Energy Storage Role: From Arbitrage Tool to Core Power Supply Guarantee
In the traditional power system, energy storage mainly serves for the short-term balance of the power grid, and its profit relies on market trading arbitrage. However, in the era of high-load electricity consumption driven by AI, the core function of energy storage has changed. It gradually transforms into providing stable and predictable electricity for high-volatile computing power loads.
On one hand, through financial means such as revenue swaps and long-term power purchase agreements, energy storage is transformed into an infrastructure asset with stable cash flow, freeing itself from the dependence on price fluctuations and becoming a long-term asset that is financiable and valuable. On the other hand, energy storage deeply integrates into the energy architecture of data centers, together with gas turbines and new energy power stations, forming a park-level power supply system, and becoming a key guarantee for the continuous operation of computing power.
Texas has successfully implemented a variety of innovative power supply models, which is the result of the combined effect of multiple objective conditions.