- Dummy loads burn up to 45% of the energy just to keep the power level stable
- Skeleton’s Graphenegpu delivers up to 40% more calculation with the same hardware
- AI workloads kill energy efficiency, graphenegpu smoothing out power pulling in real time
As artificial intelligence systems grow more demanding, many data centers have found consumption of almost twice as much the energy they technically need.
This overuse is not due to system failure or outdated hardware, it is rooted in how GPUs behave as their power needs can fluctuate drastically within seconds, from full throttle to idle.
To cope with, operators often implement dummy loads, deliberate energy players, to maintain a stable power pull – but these data centers deliberately slow down the benefit of tens of thousands of GPUs to prevent power cuts
Dummy loads mean massive waste of energy
Although this avoids damage and blackouts, it means that up to 45% of the energy is lost as the heat does not perform any useful calculation.
Skeleton technologies now claim that it can have a more effective alternative that allows GPUs to run with full capacity without overwhelming the grid.
The Estonian company developed Graphenegpu, a pointed shaving system using proprietary curved graphics super capitors.
Unlike lithium -based systems, these capacitors can respond to only 10 microseconds that absorb energy during idle periods and are emitted immediately when the GPU is loaded.
The result is, according to skeleton, the ability to maintain a uniform GPU performance without emphasizing the grid or resorting to throttling.
Their test suggests that the system can deliver up to 40% more flops – liquid dot operations per day. Second – using the same GPUs, simply by removing the performance penalties associated with thermal rating and impact instability.
“Graphenegpu delivers up to 40% more computing with the same energy footprint, while cutting both capital and operating costs by reducing grid upgrade needs, energy waste and cooling,” said Taavi Madiberk, CEO of Skeleton Technologies.
“Driven by our patented curved graph is a fundamental shift in how AI infrastructure can scale – sustainable and financially.”
The company also reports up to a 44% reduction in the power capacity that data centers must reserve from the web.
The core unit, Graphenegpu PCS 50, delivers up to 80 kW Speak power in a standard 1ou form factor, compatible with existing infrastructure and cooled with air or liquid.
It is important that it avoids lithium completely using Skeleton’s patented graphene -based material instead.
According to skeleton, this technology is tested under strict hyperscaler quality GPU workload with positive results. However, it is not independently tested for the real world achievement and durability.
The first shipping of this technology begins in Germany in June 2025. The company also has a US production site planned in early 2026.
