Every great city is built upon invisible foundations.
Rome's prosperity relied not only on its legions and codes but also on eleven aqueducts that delivered fresh water daily from dozens of kilometers away. London's rise was driven not just by its ports and finance, but by an underground pipe network that protected a million residents from epidemics. New York's skyline was shaped not merely by steel and capital, but by an ever-expanding power grid that delivered electricity to every skyscraper.
These systems operate in silence, and it is only when problems arise that people realize: the size a city can grow to depends not only on the height of the buildings above ground but also on the stability of the foundations below.
Currently, the AI industry is navigating a power desert. The power consumption of a single data center can almost drain the electricity required to run an entire city.
Gartner predicts that global data center power consumption will reach 565 terawatt-hours in 2026, a year-on-year increase of 26%. The International Energy Agency (IEA) estimates that, in an optimistic scenario, power consumption related to computing power could exceed 1,000 terawatt-hours by 2026.
The AI industry is building an unprecedented "city of computing power," but the lifeline of this city—electricity—is under pressure far beyond design expectations.
What kind of power system can support the stable operation of large-scale AIDCs and safeguard computing power? What are the underlying challenges?
Huawei's Source-Grid-Load-Storage AIDC provides some directions for our thinking.
Building the Foundation of Computing Power: Green Electricity is the Lifeline
Building a city requires many things: land, water, building materials, and labor. The same applies to the city of AI, where chips, algorithms, talent, and data are all indispensable. However, among all these resources, the most scarce, hardest to resolve, and most critical bottleneck right now is green electricity.
Large models are the skyscrapers rising from the ground, intelligent agents are the increasingly dense neighborhoods, and computing clusters are the ever-expanding industrial zones. The city is being built larger and faster, bringing prosperity, but there is a fatal prerequisite before construction begins: can the power of this land sustain a city?
In 2025, the total power consumption of national computing centers reached 170 billion kWh, accounting for 1.6% of the total electricity consumption of the whole society. The National Energy Administration estimates that during the "15th Five-Year Plan" period, the annual new power consumption for computing power nationwide will exceed 100 billion kWh, reaching approximately 800 billion kWh by 2030, accounting for about 6% of the total electricity consumption of the whole society. A GW-level AIDC consumes nearly 10 billion kWh annually, equivalent to the power consumption scale of a medium-sized city.
The ultimate limit of computing power is electricity, and energy is the foundation for the long-term development of AI. Without a reliable energy supply, no matter how massive the computing power is, it is merely a castle in the air. Under the "dual carbon" goals, this problem has become even more challenging.
China has included data centers in green power consumption monitoring, requiring that the proportion of green electricity in newly built data centers at national hub nodes must exceed 80%; Spain requires that at least 80% of the hourly power consumption for data centers over 1MW must come from newly added renewable energy, with hourly matching required; Singapore mandates in the DC-CFA2 bidding that over 50% of the power for winning operators must come from clean energy.
This means that computing power can expand, but it must be built on the foundation of green energy. Green electricity has become a hard threshold for whether an AIDC can be built, can run, and can survive.
However, the problem is not simply about "power shortage." The high-density computing power in the AI era is causing traditional data centers to face systemic challenges in space, power supply, and heat dissipation.
First, power cannot be supplied in, and heat cannot be dissipated out. Single-cabinet power exceeding 100 kW pushes traditional power supply architectures and air cooling solutions close to their physical limits. The power supply link must withstand higher currents and stricter power quality requirements, while the cooling system must comprehensively shift from air cooling to liquid cooling. This requires an architecture-level redesign.
Second, fluctuations on both the supply and demand sides make quality hard to stabilize. The intense power fluctuations of high-density computing clusters have a non-negligible impact on the power grid. Energy storage systems need to smooth out fluctuations, achieve energy time-shifting, and provide emergency backup on a time scale from milliseconds to hours, thereby improving the power supply quality and grid friendliness of AIDCs.
Third, rigid load and insufficient coordination. Computing tasks naturally possess flexible characteristics of being transferable and dispatchable, making them high-quality adjustable resources in the power system. However, traditional AIDCs are merely "power-guzzling beasts" passively waiting for power from the grid, failing to become an organic part of the new power system. Computing-power and electricity coordination, as well as source-load interaction, remain unsolved problems.
Fourth, traditional power grid construction is too slow. The iteration cycles for models and chips are measured in months or weeks, whereas power grid expansion, finding, and building new power sources often take years, making it difficult to support the rapid iteration demands of AI at this stage.
Based on these fourfold challenges, providing stable and reliable green electricity for data centers has also become one of the most fundamental and thorny problems in the AI era.
Huawei's Source-Grid-Load-Storage AIDC: Making Every Watt Produce More Tokens
With these fourfold contradictions laid out, a practical question emerges: if the power source is far away, fluctuations are the norm, and power grid construction cannot keep up, how exactly should an AIDC be built?
In the past few years, the industry's customary practice was to "build the data center first, then find the power." Following the linear process of project approval, site selection, construction, and power connection, power issues were often considered last. But in the AI era, this logic no longer works. The startup, shutdown, and load fluctuations of a large-scale AIDC park are enough to affect grid operation; with token demand growing exponentially and AIDC parks scaling from megawatts to gigawatts, the relationship between computing power and energy is shifting from going their separate ways to having to move forward together.
As the problem changes, the solution must naturally change as well. Huawei Digital Power has proposed the Source-Grid-Load-Storage AIDC solution.
The core of this system revolves around the reconstruction of three major technical fields: Watt (power electronics technology), Bit (digital technology), and Heat (thermal management technology), plus the reconstruction of the construction model, forming a "3+1" reconstruction to create the Source-Grid-Load-Storage AIDC solution. He Bo, Vice President of Huawei Digital Power, summarized its goal in one sentence: making every watt produce more tokens.
1. On the power source side, let green electricity grow on the edge of the city.
The power source is the starting point of the entire computing power city. Where the electricity comes from and how it is delivered determines whether all subsequent links can be established.
The integrated Source-Grid-Load-Storage solution enables the local consumption of new energy. In areas rich in new energy resources, wind power, photovoltaics, and AIDCs are planned synchronously to achieve local power generation from new energy and local consumption of computing loads. The core technical support for this link is Watt (power electronics technology). The industry's first 1000V grid-forming inverter FusionSolar series and the world's first 1000V intelligent string grid-forming energy storage platform LUTERRA have greatly enhanced the grid-forming capability of new energy.
2. On the grid side and energy storage side, provide a buffer for fluctuations.
The power source solves "where the electricity comes from," but the natural volatility of green electricity dictates that having a power source alone is not enough. A buffer zone is also needed during the power transmission process.
The spatiotemporal mismatch and dual fluctuations between computing power and green electricity require energy storage systems to smooth power fluctuations, achieve energy time-shifting, and provide emergency backup. Huawei's solution is to move from stable power consumption to supporting the power grid. On the source, grid, and storage sides, Huawei enhances the adaptability of AIDCs to weak grids and large load fluctuations through the coordination of grid-forming control and energy storage. The LUTERRA platform possesses six major grid-forming capabilities: short-circuit capacity support, broadband oscillation damping, inertia response, primary frequency regulation, black start, and grid-connected/off-grid switching, enabling the energy storage system to shift from grid-following to grid-forming.
3. On the load side: make computing power a flexible resource for the power grid.
The electricity is delivered, and the fluctuations are stabilized. The next question to answer is: can the largest power consumer in this city—the computing power itself—shift from passive consumption to active regulation?
Computing tasks are high-quality adjustable resources in the power system. Through grid dispatch and demand response, AIDCs can be allowed to run at full computing load during periods of high new energy generation and low nighttime electricity prices, while reducing computing power and delaying computing tasks to off-peak periods during grid peaks and tight power supply. Future AIDCs must not only adapt to weak grids and operate stably during grid fluctuations but also support the grid, becoming a grid-friendly key node.
On the load side, Huawei will build an integrated grid-forming energy router based on SST (Solid State Transformer) technology, promoting the UPS in the AIDC era from reliable power supply to grid-friendly. It is reported that Huawei has become the first manufacturer to complete broadband impedance testing for its mainstream UPS series for AIDCs. A grid-friendly UPS can achieve bidirectional friendliness between the grid and the load, supporting fluctuation smoothing, fault ride-through, oscillation suppression, and orderly recovery.
The previous three links answer the questions of "how to get the electricity, how to stabilize it, and how to regulate it," but for a city to be truly built, it must also answer "how fast it can be built."
In terms of construction mode, Huawei shifts the complexity forward to the factory through productization, prefabrication, and modularization. Taking the 18MW emergency expansion of SenseTime's Shanghai Lingang Intelligent Computing Center as an example, the delivery cycle for traditional power supply and distribution systems is as long as 3 months. By adopting Huawei's power module and lithium battery SmartLi container deployment solution, through factory prefabrication and outdoor deployment, the full-process delivery was completed in just 45 days, while simultaneously releasing all indoor space for IT use.
Letting green electricity grow on the edge of the city, providing a buffer for fluctuations, making computing power an adjustable flexible resource, and keeping the construction pace in sync with the city's expansion... the Source-Grid-Load-Storage AIDC solution integrates the power source, power grid, energy storage, load, and construction mode into a cohesive whole.
Nourishing the City with Power, Enriching the Oasis with Green
Rome's aqueducts, London's sewers... these silent systems determine how stable urban operation and management can be. The computing power city in the AI era similarly requires a power foundation that is resilient enough, green enough, and intelligent enough.
In 2026, "computing-power and electricity coordination" was written into the Government Work Report for the first time and included in the new infrastructure projects. The outline of the "15th Five-Year Plan" explicitly proposed "promoting the coordinated layout of green power and computing power."
Huawei's Source-Grid-Load-Storage AIDC is essentially designing a brand-new energy solution for the computing power desert, ensuring the stable operation of high-value computing power with high reliability, maximizing the value of every kilowatt-hour with high energy efficiency, accelerating the online deployment of computing power with fast delivery, and enabling AIDCs to support the grid while acting as a large load through grid friendliness.
The changes brought by this system are structural. The power source grows on the edge of the city, with green electricity consumed locally, no longer relying on long-distance transmission; energy storage regulates within the system, smoothing wind and solar fluctuations, making unstable green electricity adjustable and controllable; the power grid shifts from unidirectional transmission to bidirectional interaction, and computing centers are no longer just passive endpoints receiving electricity; computing loads shift from passive consumption to active regulation, becoming a dispatchable and responsive flexible force in the power system. Thus, AIDCs break free from the single role of "power-guzzling beasts" and transform into flexible resources that actively participate in power system regulation.
At the same time, the oasis can also nourish the ecology. AIDCs shift from consuming electricity to regulating it, running at full load when power is abundant and shifting to off-peak when power is tight, actively participating in grid dispatch. When computing loads can actively coordinate with the grid to smooth wind and solar fluctuations, participate in frequency and voltage regulation, and peak shaving and valley filling, the computing power city can also become a guardian of the oasis.
In short, with the entire engineering project for creating an oasis, a power desert can also be transformed into an oasis. And when green electricity is continuously injected into AIDCs, the rapid sprint of intelligence will truly have the confidence to go full throttle.