Generative artificial intelligence (AI), such as ChatGPT, DALL-E and Midjourney, is set to change the way the world works, creates and lives.

Advertising agencies are already utilising generative AI for their creative design processes, such as reference concept artwork, which has reportedly led to significant time savings. This assumes, of course, that the prompts are specific enough to yield the right results.

Microsoft is working with a healthcare software company to use GPT-4’s AI language model to draft responses from healthcare workers to patients. They will also use it to analyse medical records to identify trends.

A surge in demand for computer power

To run these generative AI apps, firms like Open AI are reportedly using Nvidia’s DGX H100 systems. If you combine nine DGX H100 systems together in a single rack, together with some third-party storage and networking equipment, you’ll get a DGX POD*. This can be used to deliver AI-based services at scale. Combining 32 of these racks would give you an exaflop* of processing power for AI and machine-learning applications.

What does this mean for real estate? These systems need power and they need space. We expect to see changes in computer architecture, network topology and data centre design because of the massive computing resources that these systems require.

The power crunch

While this is all well and good for data centre demand, the most obvious challenge is power. For a data centre currently configured for typical enterprise applications, the power requirement is usually around 7-10 kilowatts (kW) per rack. But for AI, the power requirement increases to more than 30kW per rack. And with more power comes more heat. AI applications are expected to push graphics processing units (GPUs)*, such as the DGX H100, into generating a sweltering 700 watts of thermal design power (TDP)*. This would test the cooling limits of most current data centres.

It's also becoming increasingly challenging to find shared land areas to house data centres, as they have to be able to support the high-power consumption of AI. This means we are likely to see more deals for data centres that can offer power requirements of over 100 megawatts (MW).

AI and real estate

AI is likely to have two key effects on real estate. Firstly, places with power constraints or soft limits on data centre development (because of power restrictions) may not be suitable for data centre operators that want to capitalise on AI growth.

The second aspect is ESG (environmental, social and governance) concerns. Balancing power consumption with efficient power and water usage efficiencies (known as PUE and WUE) are critical to ensure any data centre will be future-fit. Tenants are putting pressure on landlords to meet sustainability standards, where possible. Major companies that own and operate their own data centres are in an easier position. Firms like Microsoft, Apple and Google became carbon neutral by 2020 because of carbon offsets. In the UK, Amazon has supported Scottish Power’s wind farm and is purchasing its entire 50-MW output. YTL, a Malaysian infrastructure conglomerate, is also working on utilising 1500 acres of land adjacent to its 500MW data centre.

Smaller colocation providers* may have to resort to procuring renewable power from power purchase agreements. These contracts are usually expensive, costing an average of US$200 per MWh (megawatt hour). However, the providers could join forces and aggregate their purchasing power to optimise energy procurement and storage.

Development opportunities in APAC

Asia-Pacific (APAC) companies could have a pivotal hand in the AI race. Firms such as Alibaba, Huawei, Baidu and Tencent have significantly smaller self-build capacity than their US counterparts because of limitations such as economies of scale and regulatory hurdles. This presents an opportunity in APAC to provide high-density colocation data centres, which would be a viable option for large-scale tenants that are trying to increase their AI capabilities.

The use of 5G technologies in precision manufacturing or engineering, which require real-time insights at low latencies, would also spur the growth of edge data centres* located close to or within manufacturing or mining clusters. The edge data centre could then also be paired with a high-density colocation data centre that is further away, which runs the core AI models for, say, predictive maintenance. All this creates opportunities for new data centre markets.

More players will emerge as AI technology progresses. This will lead to increasingly complex AI applications and new types of data centres. It’s also clear that current global data centre capacity will not sufficiently keep up with the rapid development of AI, and this presents opportunities for investors in the data centre sector.

That said, we particularly like APAC’s prospects as we expect increasing demand for wholesale colocation* and enterprise data centres* as Asia-based technology and manufacturing firms look to scale AI capabilities in APAC. The growth of 5G applications in manufacturing activities means that there’s likely to be strong growth and an expansion of investible opportunities in data centres over the next few years.

Overall, AI is going to become more complex and more participants will want AI capabilities. We expect the potential of software to outpace current hardware capabilities in the future. The demand for data centres will increase, which will then lead to more opportunities for real estate investors.




Colocation (colo) data centres

Any large data centre facility that rents out rack space to third parties for their servers or other network equipment. Multiple tenants may occupy the facility at the same time.


Nine Nvidia DGX H100 systems combined in a rack with third-party storage and networking equipment.

Edge data centres

Smaller facilities located close to the populations they serve, which deliver cloud computing resources and cached content to end-users. They typically connect to a larger central data centre or multiple data centres.

Enterprise data centres

Data centres owned and operated by a single company.


Flops are a measure of computing power. A computer with 1000 flops can do 1000 floating-point maths operations in a second. An exaflop equals a quintillion flops.

A standard desktop computer is only able to do around 350 gigaflops.

Graphics processing unit

Designed for parallel processing, the GPU is used in a wide range of applications, including graphics and video rendering. Although they’re best known for their capabilities in gaming, GPUs are becoming more popular for use in creative production and artificial intelligence.

Thermal design power

A measure of what a computer component will generate in heat under the maximum load.

Wholesale colocation


Wholesale colocation data centres that are rented out to just a single tenant.