Summary
- VDE ITG and VDE ETG say AI training and inference create different transient load profiles that complicate grid connection.
- The briefing points to 800V DC distribution, DC/DC conversion, battery storage, and UPS integration as tools for reducing peak-load stress.
- The work gives German and European developers a more practical power architecture for dense AI facilities in constrained metropolitan grids.
VDE has set out a technical route for connecting high-performance AI data centres to constrained electricity grids, putting peak-load control, storage design, and power distribution standards at the centre of the next wave of capacity growth.
The German engineering association’s information technology and power engineering societies have published a briefing on intelligent peak-load management for high-performance data centres. The paper focuses on AI facilities where large-scale training and inference workloads create different power profiles, including transient demand swings that are harder to accommodate than steadier enterprise IT loads.
The VDE briefing says grid connection is becoming more difficult as investment in AI data centres accelerates and electricity networks face higher demands for stability and resilience. Its proposed answer is not limited to additional grid infrastructure. The paper argues that AI data centres must also change how they draw, convert, store, and manage power inside the facility.
Power distribution moves above 48V
One of the more concrete recommendations concerns internal distribution voltage. VDE says conventional 48V systems are reaching their limits in high-performance AI racks, where rising current levels increase losses and copper requirements. The paper points to 800V DC distribution as a way to reduce those losses, with DC/DC converters stepping voltage down to 48V so existing systems can still be supported.
That change reaches well beyond a single electrical design choice. Higher-voltage DC distribution affects fault protection, maintenance procedures, equipment interoperability, safety processes, and the relationship between IT hardware and the building power chain. In dense AI environments, the electrical system has to support racks that draw far more power than many legacy data halls were built to handle, while preserving uptime and maintainability.
VDE says current DC/DC converters can achieve efficiencies of up to 98%, while researchers are working to reduce the number of conversion stages. Small gains at component level become material at campus level when the IT load runs into tens or hundreds of megawatts. Conversion losses also become heat, which feeds back into cooling demand, plant sizing, operating cost, and grid draw.
Batteries take on a second role
The briefing also supports a multi-stage storage architecture, combining battery storage and UPS systems to respond across different time scales. Short-duration power quality events, training-related ramps, grid-side restrictions, and resilience requirements cannot all be handled by a single backup layer. The paper points towards storage stacks that sit between workload behaviour, UPS design, batteries, and grid operator rules.
That changes the role of the battery. Backup systems have traditionally been framed around continuity when mains power fails, but AI load management adds a second function. Batteries and power electronics can help control the facility’s interaction with the grid, smoothing peaks and reducing sudden changes in demand, while still preserving enough resilience for critical operations.
The distinction between training and inference is also useful. Training jobs can create intense and scheduled blocks of load, while inference may follow more variable user demand. A 100MW AI data centre dominated by training can behave differently from a 100MW site running inference-heavy services. Power strategy will increasingly depend on workload mix, not only on nameplate capacity.
Grid access becomes part of the design brief
Across Europe, data centre growth is being shaped by what electricity networks can accept. In Germany, the Netherlands, Ireland, the UK, and parts of the Nordics, large connection requests are colliding with transmission constraints, substation lead times, and local pressure over energy use. Developers that can show predictable ramping, internal storage, and standardised power interfaces may have an easier conversation with grid companies than sites presenting unmanaged demand growth.
Standardisation runs through the VDE paper. AI data centres are being built quickly, with server vendors, power electronics suppliers, cooling providers, and operators pushing different architectures into the market. Without common technical pathways, every large project risks becoming a bespoke integration exercise. Standardising parts of the power chain would help suppliers scale, reduce engineering risk, and give grid operators clearer assumptions about future AI loads.
The briefing also brings power architecture closer to commercial risk. Grid access is no longer a background utility issue; it is a condition of capacity delivery. Sites with land, customers, and capital can still be held back by connection terms, reinforcement delays, and load-management requirements. AI data centres that are designed to behave more predictably on the grid may gain an edge in markets where electricity capacity is already contested.
VDE’s work gives the German market a practical framework for that shift. Higher-voltage DC distribution, efficient conversion, storage layers, and standardised controls will not remove the need for grid investment, but they can change the way high-density facilities are integrated into stressed power systems.
