Grid Connection Becomes the Bottleneck for AI Data Centers

Connecting high-performance AI data centers to the power grid is becoming increasingly problematic. The VDE states this unmistakably in a joint short brief from ITG and ETG dated 16 July 2026. Operators and planners must plan load profiles, voltage levels, and on-site storage as rigorously as the compute capacity itself.

Key takeaways

Related:When the Data Center Breathes with the Power Grid  /  800 Volts in the Data Center: NVIDIA’s Push for AI

Training and inference pull differently on the grid

Classic data centers often run with a relatively steady baseload. Servers run, cooling load follows the daily pattern, and the connection is sized to a plannable average plus reserve. AI load breaks this pattern. Training and inference form two distinct load families with clearly different power and energy profiles.

Training drives racks close to rated power over long phases. Collective communication steps and checkpoint cycles create short, steep peaks above that. Inference responds more strongly to request volume. It can run more quietly for hours and then ramp up within minutes when a product launch or a batch job fills the queue. For the grid connection, what matters is less the annual kilowatt-hour than the power draw in seconds and minutes.

Damian Dudek, Managing Director of VDE ITG, draws attention to exactly this: the differences demand an analysis of energy and power demand in transient consumption. Only then can mechanisms take hold that balance load peaks. Anyone who only reads the average from the energy report is planning past the switchgear. The report must not smooth out the peak curve.

Compute capacity is orderable, connection capacity is not

The VDE sums up the situation more succinctly than many market reports: connecting high-performance AI data centers to the power grid is becoming increasingly problematic. Grid expansion is expensive, and the demands on stability and resilience grow with every large hall. Compute capacity is orderable. Connection capacity at the site often is not. Between contract signature and released capacity lie planning rounds that no GPU roadmap shortens.

Prof. Dr. Andreas Ulbig of RWTH Aachen and board member of VDE ETG classifies the grid connection as a substantial challenge. Regulatory frameworks and the availability of energy and connection capacity collide hard, especially in metropolitan areas. For operators, that means: the critical decision often falls years before the first rack, in the load forecast submitted to the grid operator.

German data centers consumed around 21.3 billion kilowatt-hours in 2025, according to Bitkom and Borderstep. Connection capacity used for AI is expected to rise from 530 megawatts to 2,020 megawatts by 2030, against a total capacity most recently of 2,980 megawatts. These figures explain why connection questions are moving from the background into the foreground. More load meets grids that do not grow in step with GPU supply chains.

I have stood in halls where the nameplate on the distribution looked clean and the measurement curve still showed spikes that no monthly average reveals. Peaks are not Excel noise. They decide whether the grid operator releases the next construction phase or sends the plan back.

48 volts hits its limits, 800 volts cuts losses

Power distribution in the building is the second bottleneck after the grid connection. Prof. Dr.-Ing. Gerd Griepentrog of TU Darmstadt makes it clear: conventional 48-volt distribution systems hit their limits at the high power levels of modern AI racks. More current at low voltage means thicker copper cross-sections, higher losses, and less room in the shaft.

Switching to an 800-volt direct current distribution system significantly reduces current losses and copper demand, according to the VDE short brief. DC/DC converters step down from 800 volts to 48 volts and keep existing 48-volt components at the rack connection-ready. According to Griepentrog, these converters currently reach efficiencies of up to 98 percent and deliver a precise output voltage. Research aims to reduce conversion stages further.

The voltage level is one lever among several. The decisive question for operators is: which voltage level do I carry from the building feed to the rack edge, and how early do I lock that in? Late redesign costs space, copper, and time. Early architecture decisions relieve the connection and the operating balance sheet. Anyone who only thinks about distribution after ordering the IT pays twice: in material and in delay.

Layer storage and factor in power density

ITG and ETG propose multi-layer storage architectures to absorb load fluctuations on different timescales. Battery storage and uninterruptible power supply (UPS) interlock: short peaks here, longer balancing windows there. Many operators are already implementing this. The concrete design depends, according to Dudek, on site, training or inference requirements, and the grid operator’s specifications.

Controlling load peaks means in practice: measure the power that the grid connection contract allows. Smooth it internally so the peak does not breach the contract. Storage buys time. It does not replace grid expansion, but it makes connections usable that would otherwise block under unconstrained transient load. Anyone who does not model peaks either buys too much connection reserve or risks curtailment in operation.

Power density per rack drives both. More kilowatts on the same floor space load distribution and cooling at the same time. Cooling is not a separate product chapter here. It is part of the power balance: every watt of waste heat must be removed and increases the building’s ancillary load. Anyone who plans only the IT kilowatts and adds cooling later systematically underestimates connection demand. In practice, IT and facility sit on the same application.

The standardization gap remains, decisions do not wait

Different voltage levels have so far been common in AI data centers. The VDE sees standardization as the key. Dudek names the benefits: size compensation equipment better, open the market for suitable technology, and accelerate the integration of renewable energy. As long as every campus runs its own voltage and storage combinations, every plan remains more expensive and harder to compare.

Operators can still act. First, analyze energy and power demand separately, including transient peaks from training and inference. Second, set the voltage level of building distribution early and coordinate it with the grid operator. Third, size the storage architecture beyond the emergency-power case, as a connection and operating instrument. Fourth, carry cooling along in power density so the connection application reflects real building load.

The short brief sells nothing. That is exactly why it is useful for decision-makers: it describes the bottleneck from an engineering perspective and names levers inside the operator’s own building. Models and benchmarks remain important. Without power in front of them, they remain slides. The operational mandate is to plan the power path and the workload together.

Frequently Asked Questions

Image source: AI-generated (July 2026)