The fast energy stores smoothing AI’s power spikes

An AI data centre can have enough electricity for its work and still create a problem when thousands of processors change gear together. Their demand rises and falls in unison. A fast energy store lets the supply connection see a gentler version of the same job.[2]

That is the market behind Siemens’ announcement on 14 October: a strategic investment and an exclusive partnership to sell and scale Skeleton Technologies’ products for data centres. Siemens, already an investor since 2023, plans to add Skeleton’s supercapacitors to its electrical infrastructure offering. The announcement leaves completion subject to the usual formalities.[1]

Why the processors move together

Training a large model involves repeated changes of activity. Processors calculate on their assigned data, then exchange the information needed to update a shared model. The arithmetic can demand much more power than the communication. A 2025 paper by researchers at Microsoft, OpenAI and NVIDIA examines production power traces and the resulting problem: synchronisation makes the fluctuations add together across the installation.[2]

The fast store charges during a quieter interval and releases energy during a burst. In the electric double-layer supercapacitors described by Eaton, charge accumulates at the interfaces between electrodes and an electrolyte. The attraction is a combination of rapid delivery and frequent cycling. Eaton describes installations ranging from individual racks to systems serving large sections of a data centre.[3]

A large burst can use little energy

Power describes the rate of delivery; energy describes the amount delivered over time. Consider a deliberately simplified example in which a buffer supplies an extra 50 kilowatts while the upstream supply stays steady. The energy required is that extra power multiplied by the length of the burst. A pulse lasting one hundredth of a second needs only 500 joules.

The figures explain why a buffer can have an impressive power rating and a modest energy capacity. Its challenge is to deliver the energy quickly, accept another charge and repeat the process. Extend the burst from ten milliseconds to one second and the energy requirement grows one hundredfold. The repeating pattern also matters: the store needs a chance to recharge.

NVIDIA provides a concrete example of this small-energy approach. Its GB300 power-supply design uses electrolytic capacitors, distinct from Skeleton’s supercapacitor offering, and specifies 65 joules of storage per GPU. For 72 GPUs, our multiplication gives 4,680 joules, or 1.3 watt-hours. That indicates scale; it does not establish usable runtime under a particular voltage range or controller setting.[4]

The company also reports a measured comparison using the same GB200 rack and Megatron training workload with two power-supply designs. Adding the new energy-storage supply reduced the peak AC demand by 30% in that test. A smaller peak does not mean the training consumed 30% less electricity: the buffer moves energy between different moments, and the reported measurement concerns peak demand.[4]

The control system earns its place

A capacitor cannot decide when to release its charge. Power electronics and control software make it a useful buffer. That division of labour is visible in Skeleton’s September agreement with Infineon: the companies intend to combine supercapacitor storage with gallium-nitride power semiconductors for peak shaving. The agreement also covers solid-state transformer development; it describes collaborative engineering, rather than measured performance from a completed installation.[6]

Eaton’s broader grid-integration paper explains why a fixed, repeating demonstration waveform is insufficient for design. Real jobs change their timing. A controller has to prevent the store becoming completely full or empty as those timings vary. Otherwise the next upward or downward movement arrives when the buffer has no room to help.[5]

The same paper treats batteries and supercapacitors as choices for different combinations of power, duration and cycling. Suitable batteries can smooth loads, although meeting rapid, repeated demand may require substantial capacity. Supercapacitors tolerate frequent cycling but hold relatively little energy. A few seconds of buffering and enough reserve to bridge a longer supply interruption are different engineering tasks.[5]

Smoothing can also carry an energy cost beyond storage losses. NVIDIA describes a GPU power-burn mechanism that temporarily maintains demand after useful work falls away, then lowers it gradually. The objective is a controlled electrical transition. It is another reason to keep a flatter power trace separate from an energy-saving claim.[4]

Siemens says the Skeleton offering can fit existing AC installations as well as provide a route towards future 800-volt DC designs. That makes retrofits part of the commercial proposition. Whether a particular site gains room for more computation depends on what actually limits it: a troublesome short peak, sustained electricity demand, cooling, or another constraint.[1]

The useful question is therefore how much of a facility’s electrical headroom is being reserved for a moment. When the bottleneck is a brief, repeated burst, a fast store can change what the upstream equipment must deliver. It does that by lending the computation electricity for a very short time.

Report an error or suggest a correction