Skip to content

Neoclouds and power constraints: why Power Compute Effectiveness matters

13 November 20262 min read
Guest Insights
Glowing digital cloud icon connected by data nodes on a dark blue background

Access to GPUs and accelerators is no longer the only bottleneck facing AI infrastructure. For operators building capacity in the UK and more widely, the harder question is whether there is sufficient usable power, delivered to the right site at the right time, to run the hardware. Grid connections and planning timelines are long while compute demand continues unabated.  

Specialist GPU providers – also referred to as neoclouds – such as CoreWeave, Nebius, Nscale, Crusoe and Lambda are scaling quickly to fulfil the GPU demand from hyperscale and AI lab customers. Analyst group Gartner expects neocloud providers to capture 20% of the $267 billion AI cloud market by 2030 while Synergy Research Group forecasts that the neocloud market could approach $400 billion by 2031. 

However, the growth of neoclouds – and the UK’s data centre capacity more generally – is hitting significant power-related speed bumps. The UK’s transmission and distribution grid backlog means data centre projects face multi-year connection queues. For example, Microsoft recently announced its decision to pull out of a planned data centre project in the UK with leading neocloud Nscale after it emerged the site in Loughton, Essex, might not receive power until 2030. 

Given that kind of power-related pressure on new builds, the impetus to ensure that power is used as effectively as possible on every new project is intensifying. A data centre’s headline power capacity is not the same as its compute capacity. As well as the actual GPUs and compute, provisioned capacity also goes to cooling, power conversion and other supporting infrastructure. Legacy data centre designs can lose 15% to 25% of provisioned power to overhead architectures.  

Power Usage Effectiveness (PUE) has long been the standard measure of those non-compute losses. It compares total facility power with IT equipment power and asks how much energy is lost to overhead. Lower is better, and 1.0 is the theoretical ideal. PUE answers an operational question: how well is this facility running? It is only useful after the site has been designed, built and put into operation. That approach was workable when power availability was not so constrained – PUE could be used to guide an operational facility in the right direction. 

However, given an increasingly power-constrained backdrop, a different metric is required to optimise data centre design and build. Power Compute Effectiveness (PCE) asks a different question to PUE: how much of the power a site is designed to provide is available for compute? It is calculated as: 

PCE=Provisioned IT PowerProvisioned Total Facility Power AllocationPCE=Provisioned IT PowerProvisioned Total Facility Power Allocation 

Expressed as a value between 0 and 1, PCE was introduced in a white paper published by the Open Compute Project (OCP) and co-authored by data centre infrastructure specialist Airsys. 

The two metrics work together. A facility can have excellent PUE, with very efficient cooling, and still allocate a small share of its provisioned power to compute because its design reserves too much structural capacity for support systems. Operational efficiency and structural allocation are separate properties, and AI infrastructure planning needs both. PCE is also a planning metric based on provisioned rather than measured power, so it is most useful at the design and procurement stage. 

What good looks like 

Facilities that convert power into compute effectively – including many neocloud sites – tend to share a set of common design features: 

  • High-density, liquid-ready designs. Rack densities rise with each GPU generation. Infrastructure built for liquid from the outset avoids redesigning the facility each time. 

  • Closed-loop cooling. Systems that recirculate coolant through chillers and dry coolers keep water use low. Coolant loss in such systems is minimal, so attention is needed only every few years. 

  • Heat reuse where it is viable. Nebius’s Mäntsälä site in Finland recovers around 20,000 MWh of heat a year, enough to heat about 2,500 Finnish homes. Heat reuse does not depend on one cooling technology, but it must be designed in from the start.  

  • Flexible air and liquid deployment. Many sites will run both, across new builds and retrofits. 

  • Power allocation designed for compute. The share of provisioned power available to compute should be a design target, not an afterthought. 

Meeting demand for cloud and AI will require not only new clean capacity but also higher compute effectiveness from existing and planned capacity. Where grid connections are scarce and slow, every megawatt that is converted into compute is capacity added without waiting for a new connection. Ultimately, the successful operators of the future will need to convert more of every megawatt of power into usable AI compute, and metrics like PCE will help guide them towards that goal.