6 key takeaways from Data Centre World 2024

Download Listen

We were at Data Centre World 2024 in London last week to see what the world’s foremost data centre providers are focusing on today. This article gives our 6 key takeaways from the event.

1. AI is key driver of demand for new data centre capacity

450TWh of demand for data centres in 2023 will grow to 800 TWh by 2026 according to many estimates (equal to roughly 8% of global electricity consumption). This is largely driven by the expected increase in demand for AI applications. AI applications cannot be run on much existing infrastructure as AI development/training requires new server components and new architectures (e.g. server ‘clusters’ with a high number of GPUs). Future sites are likely to be hybrids, able to run both AI and legacy workloads. While applications such as autonomous vehicles have been driving the hype as opposed to material demand, other applications such as computer vision are likely to require capacity in the short term.

2. Retrofitting brownfield sites to host AI workloads presents challenges

150 racks previously required about 800KW of power. This would only power about 8 racks if training AI workloads. This is due to the need for server clusters filled with power-hungry GPUs. This means that brownfield sites would likely need to secure much more power to train AI workloads. Even if you have the power, retrofitting brownfield sites presents significant complexity in upgrading/replacing cooling solutions. Compressed AI models have lower power requirements once deployed, but where you need to integrate new AI servers into existing sites there will be challenges.

3. Difficulty securing power is holding back greenfield build out

It many markets there is a big lag in connecting greenfield sites to the grid, and regulators/governments like Ofgem in the UK are often accused of having highly ineffective processes for dealing with connection requests. Moreover, in some key markets there is even a moratorium on connecting new facilities to the grid (Dublin, Singapore). Energy prices remain high in much of the world which creates additional pain for facilities. Another issue facing greenfield builds is supply chain constraints, as geopolitical events put pressure on already stretched supply chains of key components like chipsets (GPUs in particular). Large new manufacturing facilities from the likes of NVIDIA are on the way but these pressures will continue to exist in the short term.

4.  AI will need the edge, not just for latency, but to distribute capacity

Up to 50% of AI workloads may be at the edge by 2028. While development/training workloads and large language models (LLMs) will stay in hyperscale facilities, we expect regional data centres to be hosting AI workloads (compressed inference as opposed to development). While it remains difficult to build new sites in FLAP-D and other major locations, the answer may be to move some of this capacity to the edge. Applications such as computer vision which we expect to drive early AI demand at the edge also have the latency requirements which makes distribution closer to end users highly valuable.

5. Data centres must find new power sources to reduce reliance on inadequate and expensive national grids

In order to solve the challenge around access to reliable, inexpensive power, many facilities will include on-site power sources. Sites being able to generate on-site solar, wind or other renewables has long been touted as a possibility, but other solutions such as hydrogen fuel cells are beginning to gather traction. However, this is likely to be as back-up power or as part of a mix as there is difficulty scaling. There are also reliability issues if facilities are to rely completely on on-site generation. An exciting area of development is on the nuclear side, where small modular reactors have capacity of up to 300MW.

6. Data centres can earn revenue by supporting legacy grid systems

Demand (and with renewable energy, supply) on grid systems has peaks and troughs. In order to match this demand and supply, grids wish to store energy in times of low demand and then deploy it when demand rises. This ‘load balancing’ is becoming a greater challenge as we produce more of our energy from renewable sources which are intermittent. Data centre UPS systems can sell power back to the grid when demand is high, and act as a storage for excess energy when demand is low. Enabling more effective load balancing is seen as a vital element of future grid systems and would be very valuable. We are already seeing some data centre infrastructure providers, such as Vertiv, working with data centre operators to bring this vision to life.