Huawei Unveils Source-Grid-Load-Storage AIDC 1.0 as AI Data Centers Face Power Bottlenecks

The timing reflects a broader shift. On Sept. 16, Nvidia, Google and Emerald AI announced the AI Energy Management Alliance, or AEMA, to push data centers to adjust electricity use dynamically according to grid conditions and become more flexible grid resources. The International Energy Agency warned in a 2025 report that about 20% of planned data center projects could be delayed if grid-related risks are not resolved, and in a 2026 report projected global data center electricity consumption to rise from 485 TWh in 2025 to 950 TWh in 2030, with AI-oriented data center demand nearly tripling over the same period, according to QbitAI.

Policy is moving in the same direction. Spain in August 2026 published a draft royal decree that would impose stricter renewable requirements on new grid-connected data centers of at least 1 MW. Each operating hour would need at least 80% of electricity matched by newly added renewable generation in the same hour, and every 1 MW of new power demand would require 1 MW of new renewable capacity built 18 months before the data center begins operation, either self-built or through long-term power purchase agreements. Noncompliant projects could face surcharges, higher network rates or, in cases of persistent violation, loss of grid access. China in April 2026 issued an action plan on the two-way empowerment of AI and energy, published in May with 29 tasks, encouraging grid-forming storage and green power direct connections at computing facilities. The European Commission adopted a unified data center sustainability rating scheme on Sept. 21, 2026, with first labels expected in 2027.

Inside the data center, the problem is not only total capacity. AI clusters can move from computing to communication to idle states in sync, creating fast power fluctuations that stress supply systems. Two data centers with the same annual consumption may have very different needs if one has a smooth load and the other swings sharply. The source-grid-load-storage approach is meant to consider generation, transmission, loads and storage together. Google disclosed in March that it had included 1 GW of demand response capacity in long-term energy contracts with U.S. utilities, shifting or limiting some machine learning tasks to support the grid. Such flexibility must respect business constraints: real-time inference has latency limits and training jobs have deadlines, so the movable load depends on utilization, power prices and storage costs.

Density adds another constraint. At the same voltage, higher device power means higher current, increasing pressure on cables, connectors and distribution space. Raising supply voltage and using direct-current architectures can reduce conversion losses, and 800V DC has drawn industry attention. Nvidia, Google, Microsoft and others are advancing an 800V DC standard under the Open Compute Project, with more than 80 equipment and infrastructure companies involved, according to QbitAI.

Huawei's strategy, first outlined in May 2026 as Source-Grid-Load-Storage AIDC, groups the reconfiguration into a "3+1" framework: Watt for the power chain, Heat for thermal management, Bit for digital operations, and construction model. At the September summit, Huawei Digital Power Vice President He Bo described three stages in the data center's relationship with the grid: adapting to the grid, supporting the grid and actively forming the grid. In the first stage, a data center is mainly a load with UPS and backup power. In the second, storage shifts from backup alone to backup plus regulation, allowing peak shaving, demand control, short-term fluctuation smoothing and demand response. In the third, grid-forming storage, energy routers and microgrid controls strengthen voltage and frequency support and can support island operation under the right conditions. Huawei said this matters most for high-density campuses, green power direct-supply projects and areas with weak grids.

For the Watt layer, Huawei's AIDC 1.0 uses grid-friendly UPS, smart lithium batteries and grid-forming storage to maintain supply continuity, smooth AI load swings and support the grid. The company proposed a future MIMO power architecture with multiple energy inputs, multiple output formats and multi-level storage coordination, and a longer-term path toward a solid-state-transformer-based grid-forming energy router. Huawei said its Taishan UPS has a single-cabinet capacity of 1,280 kVA and double-conversion efficiency of up to 98%, while its Hengshan DC UPS uses one hardware platform to support 270V, 400V and 800V DC formats. It also said SmartLi 5.0 improves storage density, load smoothing and safety, and is evolving for 800V high-voltage DC.

For the Heat layer, Huawei introduced an AI-enabled liquid cooling system with a thermal management unit, or TMU, as its control core. The system adds micro leak detection, coolant health prediction and predictive maintenance, addressing the reliability needs of large-scale liquid cooling. Huawei said the TMU is natively adapted to Ascend servers and can coordinate cooling with compute loads through cold-power linkage. For the Bit layer, Huawei is applying digital and intelligent operations across design, delivery and operations, with predictive maintenance as one example. For construction, Huawei said Power Module 5.0 reduces footprint by 40% and cuts delivery time from seven days to three days. Its Power POD, introduced in May, supports 3.2 MW per box, one box per power route, plug-and-play and outdoor deployment, while its IT POD uses eight boxes as one system for a productized cooling chain.

Huawei pointed to a SenseTime Lingang AI computing center case as an example. Facing an 18 MW expansion, the two sides used power modules and SmartLi container deployment, relying on factory prefabrication and outdoor deployment to complete power distribution system delivery in 45 days, compared with about three months for a traditional comparable system. Deploying equipment on the roof also freed indoor space previously used for power distribution. Huawei also showed a 3D data center concept that borrows the separation of power and production layers from chip factories, arranging cooling, IT, power and backup systems vertically to rethink the relationship between infrastructure and computing equipment.