As AI loads swing faster and grid connections get harder, will data centres need to generate, control and stabilise their own power.
The biggest constraint on AI data centre growth may not be GPUs, land or even cooling – it could be electricity and Piller is targeting a problem that could become central to AI infrastructure: how do you keep power stable when the data centre is no longer relying entirely on the grid?
As AI campuses hurtle towards hundreds of megawatts and gigawatts, securing grid connections quickly enough is becoming increasingly difficult which is pushing developers towards a different model: generate the power on site and build the data centre around it. On paper, this solves one problem – in reality, it creates another.
The power plant now has to follow the AI
A conventional grid connection provides something that is easy to overlook: an enormous electrical buffer, but when demand suddenly changes, the grid absorbs much of that disruption – an isolated AI campus does not have the same luxury. AI workloads can produce rapid, repeated multi-megawatt changes in demand, with load moving dramatically within sub-second timeframes and whilst engines and turbines can generate huge amounts of electricity, they cannot mechanically respond to those changes at the same speed creating a potentially awkward mismatch : The data centre operates at GPU speed but the power plant operates at engine speed – so something has to sit between the two.
Piller thinks that something is power stabilisation
This is where Piller’s SHIELDX technology becomes interesting – rather than adding more generation capacity, the system is designed to absorb or inject active power during sudden changes in demand, helping to shield the generation plant from the data centre and the data centre from disturbances originating in the power plant. This idea is important because the obvious answer to unpredictable AI loads would be to oversize the generation plant which works — but it is hardly elegant as building additional engines or turbines simply to cope with short-lived peaks means more capital expenditure, more equipment and potentially lower utilisation. It is worth noting that Piller’s proposition argues better control could be more valuable than simply adding more hardware.
Will AI turn the data centre into its own power system
Pillers recent work suggests this is becoming more than a theoretical problem – it’s wider Power Solutions business has recenty secured more than 3GW of behind-the-meter power deals for AI data centres in the US, while its pipeline runs into the tens of gigawatts and Bergen Engines – part of the same business – is also supplying more than 500MW of on-site generation for AI data-centre developments that can operate in island mode or alongside the grid.
These numbers matter, but the underlying change matters more; is AI turning some data centres from electricity consumers into miniature power systems and forcing data centre operators to consider generation, grid interaction, transient loads, frequency, voltage, fault response, what happens when an engine trips and to think like power systems operators alongside their usual concerns around GPUs and cooling.
Could the next bottleneck be stability?
This is where Piller’s strategy is worth watching because they are not simply selling another UPS into an AI data hall – they are targeting the increasingly blurred boundary between the power plant and the data centre and that boundary is likely to become more important as developers look for ways around grid constraints.
The underlying infrastructure problem is real: AI loads are changing faster than conventional generation can comfortably respond and whilst the industry is focused on where the next facility will get it’s power from, what happens after they get it, because in a gigawatt-scale AI facility having enough electricity may be only the beginning – keeping it stable could be the next battleground.
