Verne Global CEO Dominic Ward speaks about the increased demand for data, the energy impact of generative AI and the importance of choosing the right location for data centres.
It’s no secret that data centres are becoming a bigger strain on energy grids as the world becomes more digitalised.
Their impact has become clear in Ireland, with 18pc of the country’s metered electricity being consumed by data centres last year, according to figures from the Central Statistics Office. The same amount of electricity was consumed by all urban dwellings in Ireland.
Despite the strain on the grid, the desire for more data centres has not gone away. Just last month, Amazon Web Services received approval to build three new data centres in Dublin, despite opposition from environmental groups.
There are clear benefits from data centres, with their ability to store vast amounts of easily-accessible data being vital for various sectors – including research. They are also highly valuable for AI, a technology that has grown significantly in recent years.
Dominic Ward, the CEO of data centre provider Verne Global, told SiliconRepublic.com that power needs in Europe have grown increasingly stretched and that this has been driven “in no small part by the growing demand for AI”.
“Dell’Oro Group predicts that global data centre capital expenditure is set to rise above $500bn by 2027 due to AI infrastructure demand,” Ward said. “The growing demand for AI, alongside the worsening climate crisis, has led to a need for data centres that are specifically designed to handle the heavy compute required by all types of AI.
“At the same time, there’s an imperative to prioritise energy efficiency so that valuable power is not wasted, and the impact on the environment is kept to a minimum.”
The demand of generative AI
When speaking about the impact AI is having on the demand for data centres, Ward said there’s an important difference between traditional AI and generative AI. The latter is the one that has gained rapid attention over the past year, with the rise of certain systems like ChatGPT and a heavier focus by big tech companies.
“Generative AI is more complex and therefore requires more computing power,” Ward said. “As an example, researchers in 2019 found that creating the generative AI model BERT consumed energy equivalent to a round-trip transcontinental flight taken by one person.”
The energy requirements for these systems has grown exponentially as they have become more complex. While Google’s BERT had roughly 345m parameters, OpenAI’s ChatGPT has 1.5bn. The company also claims that its latest large language model, GPT-4, has more than 1.75trn parameters.
“Most of the major generative AI models are generated by hyperscale cloud providers with thousands of servers,” Ward said. “These models run on graphics processing unit (GPU) chips that require 10-15 times the energy a traditional CPU needs.
“This is because a GPU uses more transistors in the arithmetic logic units that are essential for executing the calculations and operations required to run AI and machine learning models.”
Location, location, location
Despite the intense amount of energy these machines consume, it looks like the demand for generative AI systems will grow in the coming years. This will likely raise the energy demand on data centres, which will also lead to more carbon emissions.
As a result of growing demands, Ward said many data centre operators are focusing on optimisation, by ensuring that their data centres can handle the computational load of AI “as efficiently as possible” and by shifting to the use of renewable energy sources.
“This is exactly why Verne Global founded its operations in Iceland 15 years ago – with its year-round temperate climate and access to renewable energy sources, it was the perfect location to set up a data centre that would provide maximum output with limited impact on the environment,” Ward said.
“Verne also chose to build its Icelandic data centre on a former NATO base – and this brownfield approach is another example of how to reduce data centres’ impact on the environment.”
Ward spoke to SiliconRepublic.com earlier this year about the need for good digital infrastructure to make the sector more sustainable.
Some examples of measures the sector is taking to become more efficient include as alternative cooling methods like liquid cooling technologies – to boost their energy efficiency and reduce their environmental impact.
But Ward said the location of data centres is the key way to “really mitigate” the impact these locations have on the environment. He claimed that only a small percentage of applications are “truly latency-sensitive”, which means they need to be located near a data centre.
For most – including many AI applications – Ward said its possible to locate their data centres in locations that have renewable energy sources – like Iceland.
“The temperate climate of Iceland and other Nordic countries moreover means that data centres located there can utilise free air cooling – further improving energy efficiency and thereby reducing the environmental impact of the data centre and AI compute,” Ward said.
Regardless of the measures taken, it appears that more efforts will be required by data centre operators to make the sector more sustainable.
“The more complex the technology, the greater the energy needed in the data centre,” Ward said. “That’s why it’s more important than ever to prioritise data centres that have both the capacity to deal with the high intensity compute demanded by AI technology, whilst also having access to reliable and sustainable energy sources.
“If organisations focus on placing only the small percentage of applications that are truly latency-sensitive in metro locations, whilst placing the majority of their compute in data centres in optimised locations such as Iceland, then the rising demands on the data centre industry, driven largely by AI, will be far more manageable.”
At the end of 2022, the COO of digital infrastructure company Vertiv – Giordano Albertazzi – predicted that increased regulation in the data centre sector is “inevitable” due to the amount of energy the sector consumes.