CAN THE GRID KEEP UP WITH AI DATA CENTRE POWER DEMAND?

Article contents

Key Insights:

• AI data centre power demand is rising far faster than grid infrastructure can be expanded, creating a growing gap between project timelines and connection availability.
• A 500-rack traditional colocation hall operating at an average of 8 kW per rack may require around 4 MW of IT power. The same number of AI racks averaging 65 kW would require approximately 32.5 MW.
• Reducing energy consumption means improving the amount of compute delivered for each unit of power, through measures such as efficient cooling, better electrical distribution and stronger IT utilisation.
• Phased deployment, demand response and battery storage can help manage when and how AI loads reach the grid, but most large developments still depend on a reliable, high-capacity power connection.
• A power-led design strategy brings grid engagement, capacity phasing and energy modelling into the project before site and design assumptions become fixed.

AI has changed the scale of data centre power consumption

Across the UK, Europe and North America, developers are securing sites, progressing designs and committing capital to AI data centre projects, only to find the grid connection they depend on is years away - often three, four, even five years out. 

That raises a basic question: how are AI data centres powered when the connection they need isn’t yet available at full capacity? 

This is quickly becoming one of the biggest delivery challenges in AI infrastructure. Data centres that once required 10 to 20 MW are now being proposed at 100 MW and beyond, and the hardware inside them is changing at a pace that electricity networks were never designed to match. 

What you end up with is a simple but critical mismatch between how fast AI infrastructure can be planned and built and how fast grid capacity can actually be delivered. And that disconnect is now shaping everything - from where sites are viable, to when they can go live, to how much risk is sitting with developers and investors long before a single server is switched on.

This article looks at why that gap exists, what it means in practice for project delivery, and how a more power-led approach to design and site strategy can help make AI data centres work within the constraints of the grid as it actually is.

Why AI data centre power demand is rising so quickly

In a traditional data centre, the power load is spread relatively evenly. You’ve got racks of general-purpose servers, each drawing a steady, predictable load. It’s a model the industry knows well, and the supporting electrical and mechanical systems are designed around it.

AI facilities work differently - as we’ve explored in our comparison of AI data centres vs traditional data centres.

Instead of a balanced spread of load, power gets pulled into dense GPU clusters. These racks draw far higher loads than traditional IT systems were ever designed to handle, and the move from CPU-led to GPU-heavy compute is a key reason AI data centre electricity consumption is climbing so quickly. It changes the scale of everything in the building.

A typical colocation data centre

One of the simplest ways to understand the impact of AI on power demand is to start with a conventional colocation data centre.

In most multi-tenant sites, rack densities have historically stayed below 10 kW. That’s the baseline level most existing halls were designed around. The Uptime Institute’s 2024 Global Data Centre Survey put average rack densities below 8 kW, with most facilities having no racks above 30 kW.

Its 2025 update reported a gradual shift towards racks in the 10-30 kW range, but also noted that few facilities exceed 30 kW and that extreme densities remain relatively limited in real-world deployments.

The result is an environment where power distribution and air‑based cooling have developed steadily over time, built on predictable, well‑understood loads.

A GPU-dense AI data centre

AI hardware is pushing those power levels into a completely different range. Racks built around high-end training systems are now routinely designed for tens of kilowatts each, with some configurations exceeding 100 kW. 

H100-class systems (based on NVIDIA’s H100 GPU, one of the most widely used chips for AI training) are typically deployed in configurations that push rack power above 40 kW, while NVIDIA’s DGX GB200 SuperPOD comes in at around 120 kW per rack - with a 32‑rack SuperPOD drawing roughly 3.8 MW in total.

Not every AI deployment will run at those extremes, but the pattern is hard to ignore: significantly more power‑dense IT, concentrated into fewer racks, generating heat loads that traditional air cooling alone is increasingly ill‑equipped to manage. 

Every kilowatt added at the rack level has to be moved and rejected as heat, so cooling, power distribution, UPS and other supporting systems all scale alongside the IT load.

Same number of racks, very different outcome

The easiest way to visualise the impact is to keep the rack count the same and look at what happens as the average load per rack increases.

As the examples above show, a 65 kW AI rack is by no means the upper limit of what's being deployed today. That gap tells you everything about how fast AI data centre electricity consumption has moved - and how unprepared most existing sites are for it.

Some of the latest designs are already operating at around 120 kW per rack. That means a hall with the same footprint and the same number of racks can end up needing a completely different level of power once it’s configured for GPU-heavy AI workloads.

Why can't the grid expand at the same pace as data centre construction

Grid infrastructure wasn't built for this rate of growth, and it can't be expanded at the pace AI is moving.

Grid development runs on a very different timeline to data centre construction. 

A well-executed data centre build might take 18 to 24 months from investment decision to commissioning. In comparison, new substations, transmission upgrades or major reinforcements to existing networks often take five to ten years - sometimes longer once planning, consenting and supply chain pressures are factored in.

In regions where data centre demand has concentrated, including parts of the UK, Ireland, the Netherlands and Northern Virginia, grid connection queues have become much longer. Developers are quickly discovering that available capacity on paper doesn’t always mean capacity that can be delivered in the timeframe a project requires.

There may be enough energy on the grid, but can it actually deliver the right capacity where and when AI data centres need it? For most large-scale AI data centres, the grid connection still does the heavy lifting. Battery storage, backup generation and on-site energy sources can help keep things running and balance demand, but they rarely replace the need for a reliable, high-capacity grid connection.

What grid constraints mean for project delivery

This is where the mismatch between grid timelines and data centre timelines turns into a real delivery issue.

Take a developer planning a 60 MW AI campus. The site is secured, planning is in place, procurement is underway, and construction is being prepared. Then the grid offer arrives: 30 MW available upfront, with the rest delayed by around 36 months.  This is becoming increasingly common in markets where demand has moved faster than network upgrades.

• At that point, the developer's options are all compromises:
• Phase the IT rollout and carry the cost of land and development while waiting for additional capacity.
• Rework the electrical strategy to make the most of the initial connection. 
  Or step back and reassess the site altogether.
  

None of those options are cost-free, and earlier grid engagement could have allowed them to be planned for much sooner.

It also has a knock-on effect across delivery. Construction sequencing, procurement strategy and commissioning plans are all built around an assumed power-on date, and when that date moves, or when it arrives with less capacity than expected, everything built around it has to move too.

The later grid reality is factored in, the greater the risk that the whole delivery plan has to be reworked.

How to reduce energy consumption in AI data centres

One way to deal with high power demand is to use less energy for each unit of computing. That means building efficiency into the design of the facility from the start.

Cooling design is where a lot of the efficiency gains are. At the power densities AI hardware is running at, air cooling starts to struggle. Liquid cooling changes that because it moves heat out more effectively and cuts down the energy needed to deal with it. In some cases, simply letting equipment run at higher temperatures (where it’s safe to do so) can reduce cooling demand further.

Power distribution matters too, and the savings build up across the entire electrical chain. Every time power is converted, you lose a bit of efficiency, so the cleaner the electrical setup, the better the overall performance. Higher-voltage distribution helps reduce those losses, and keeping equipment properly sized (rather than oversized - “just in case”) helps systems run closer to their efficient range. The real benefit comes when electrical and cooling design are aligned from the start.

PUE is still useful, just not the full picture. The best sites are now getting under 1.2, so less than 20% overhead on things like cooling and support systems. But it doesn’t tell you how well the IT is actually being used, or whether racks are properly loaded.

That’s where IT utilisation comes in. If workloads are spread across too many idle servers, energy gets wasted. Tightening that up - running fewer active machines and matching compute more closely to available capacity - brings energy use per unit of output down.

Managing how AI power demand reaches the grid

Managing AI power demand is really about when and how the load hits the grid, not just how much power is needed. That’s becoming a bigger issue as grid capacity is stretched in a lot of regions.

Phased deployments are the most common way of dealing with it. Instead of bringing a whole 60 or 100 MW campus online at once, you bring it in stages - hall by hall, or cluster by cluster. That means the initial grid connection can be smaller, and you grow into it as more capacity becomes available. It only really works if the design is set up for it from day one, but when it is, it takes a lot of pressure off the early stages of a project.

Then you’ve got demand response, which is increasing in popularity. Some AI workloads (especially training runs and batch inference) don’t need to run at a fixed time, so you can schedule them away from peak grid periods or adjust them when the grid is under strain. There’s been early work in the UK showing that certain AI compute loads can actually modify their demand in response to grid conditions without affecting core operations. It’s not universal though - anything real-time or latency-sensitive doesn’t have that flexibility. But where it does work, it gives the grid some breathing space.

There’s also battery storage. It helps smooth out peaks and deal with short spikes in demand that the grid connection might struggle with. It won’t replace the need for grid capacity on sustained loads, but it does help reduce how much peak capacity you actually need from the connection.

Can local generation close the gap?

We hear a lot about on-site generation helping bridge the gap - things like solar, wind, batteries, fuel cells, even private-wire connections from nearby assets. And they do have a role. They can bring down annual carbon emissions, help smooth peaks, and give you a bit more flexibility while you’re waiting on grid timelines. We’ve explored this more in our roadmap to zero-carbon, water-negative data centres. 

But they don’t replace the grid for anything at AI scale.

Solar and wind are intermittent by nature. Batteries are useful, but they’re not going to carry multi-megawatt GPU clusters through long periods of sustained load on their own. Fuel cells and local generation can support the site, but they still don’t change the fundamental requirement for a strong, continuous grid connection.

So what you end up with is this: local generation can strengthen the setup and give you more options, but it doesn’t remove the underlying dependency on a solid grid connection.

Why AI data centres need a power-led design strategy

All of this really points in one direction: power can’t be something you sort out after the site’s been secured. It needs to be part of the design thinking from the start.

That means:

• Talking to network operators early, before capacity assumptions are locked in, to understand what is actually deliverable and when
• Planning demand in phases so electrical systems, cooling, and IT rollout match what the grid can support at each stage, rather than designing for a full steady-state load that isn’t available from day one
• Designing electrical and cooling systems together, especially for liquid-cooled AI deployments, so you don’t end up with infrastructure sized for capacity that isn’t available yet
• Avoiding the mismatch between installed capacity and actual utilisation, which is where a lot of waste and operational friction shows up later
  

Resilience, carbon performance and long-term operating cost all come back to those early decisions. Get the power strategy right at the start, and the rest of the project stays much more stable as it moves through delivery.

So, can the grid keep up?

Short answer? The grid can absolutely support the growth of AI data centres. But only if we stop expecting it to behave like a fixed, on-demand utility that always delivers full capacity, anywhere, on project timescales.

That’s where projects start to run into trouble.

The projects we opened with aren’t really warning stories about AI; they’re more reminders that power has to be factored in from day one, not confirmed later in the process.

The strongest developments treat power as a real design input early on, cutting unnecessary demand where they can, and thinking about capacity in stages rather than as a single step.

AI data centres and the grid can work together - they just need to be planned around each other, not treated separately.

At RED, we help developers, operators and investors work through that gap between AI ambition and grid reality - from early-stage power strategy and site selection through to phased electrical design and energy modelling for AI data centre environments. 

If you’re dealing with these questions, we’re always up for the conversation. Contact us today.