Connect with us

Capital Markets

Fintech’s data centers face a GenAI “energy paradox”

The market for data centers is booming, but artificial-intelligence use cases are an energy nightmare.



Data centers make the world go round. These buildings house the servers that store and compute every digital function our societies carry out. The “cloud” is very much grounded in racks of computers snuggled together, which in turn rely on energy, cooling systems, and connectivity, which runs the gamut of switchgears and wires to subsea cables to satellite broadcasts.

Generative artificial intelligence is piling on demand for data centers but it also creates a huge challenge. Large-language models such as ChatGPT rely on Nvidia graphic processing units to train AI models, and these chips are insatiable, energy-hungry beasts.

“Generative AI isn’t just about adding more racks,” said Steven Lee, director at Schneider Electric, a specialist in digital automation and energy management who spoke at a recent conference. “It’s changing the paradigm for data centers.”

From finance to fintech

Data centers are the unglamourous backbone to fintech and finance. The biggest ecosystems are in the US (northern Virginia and Portland, Oregon), but so are Singapore and Hong Kong, thanks to their roles as global financial centers.

Financial institutions have been the biggest customers of data centers in these Asian hubs, along with telecommunication companies.

That complexion is now changing, however, as new customers emerge competing for racks and resources. The industry can’t build enough data centers quickly enough.

The AI wave is already crashing on the America, and it will ripple its way to Asia’s data centers over the next two years, says Andrew Green of JLL, a commercial-property consultancy.

The new set of big players are hyperscalers, a term that includes large cloud providers and internet companies such as AWS, Meta or TikTok. Their hunger for computing is forcing data centers to become bigger, more flexible, and more bespoke.

New data centers required

This is changing the design requirements for data centers. Banks and insurers have been content with traditional co-location data centers, which are cheaper because they offer a standardized infrastructure. Hyperscale is required for processing-intensive work and a broader set of applications, including digital content, social media, video streaming, e-commerce, cryptocurrency, and mobile payments.

And now add genAI to the list, which is really fuel on the fire. Lee says a ChatGPT query consumes 10 times more energy than a search on Google. Ten times!

That’s on top of a broader shift to digitalization across the world, with Covid a major accelerant as work and life went remote. One example: the number of daily users of Microsoft Teams went from 20 million in 2019 to 300 million in 2023, says Damon Lim of datacenterHawk, a real-estate research firm focused on data centers.

By 2030 there will be 7.5 billion people using the internet, or about 90 percent of the global population. That demand is why there will likely be more than 1,000 hyperscale data centers in operation by the end of this year.

Paradox or power-play?

The upshot is a fourfold increase in power consumption by data centers between now and 2028, says Lee. “This is the energy paradox of huge consumption,” he said. 

But there’s nothing paradoxical here: as society digitalizes, it is finding the energy costs to feed that trend are rising exponentially, so ‘greening’ data centers is vital. And the trend is global. For example, AI use now accounts for about 8 percent of data-center consumption in Southeast Asia. That figure is set to more than double by 2028.

The trend is also spreading, which makes it harder to make sustainable. If it was just about data centers, the green focus could go on new builds or refitting existing buildings. Indeed, almost all AI-related workloads today are stored and computed centrally in data centers. But increasingly these tasks are being done at the ‘edge’, distributing them close to the source of data (such as buildings or machines, or perhaps your phone). Put AI on top, and the power requirement keeps growing – and spreading.

Data center developers are responding, in theory, by making racks of servers more energy-efficient, changing the means of cooling the computers (a particular challenge for data centers in hot, tropical locations), and plugging into electricity sources that rely more on renewables.

What sounds good on paper is difficult and expensive to carry out. Data centers haven’t been designed with these needs in mind.

Rosanna Tang, a researcher at Cushman & Wakefield, a property consultancy, notes that in Hong Kong, 44 percent of data centers are housed in old industrial buildings.

These centers can’t handle heavy-duty processing. They might only be able to consume 10 to 15 kilowatts per rack to meet a compute need, whereas hyperscalers’ needs typically eat up 40 kw/rack, and some need up to 100kw/rack.

“You’ll need dedicated infrastructure to provide that level of cooling and power support,” said Patrick McCreary of Yondr Group, a hyperscale operator and developer.

Hong Kong’s special case

In Hong Kong, that means building more centers in new areas. Most of the city’s data centers are in Kowloon, especially around a hub at Tseung Kwan O. But the promise of the Greater Bay Area and integrating with neighboring mainland cities means the new generation of data centers will be built along the Shenzhen border, in what the local government brands the ‘Northern Metropolis’.

Data centers are uniquely problematic in Hong Kong because of lack of available land and extraordinary rents, combined with stringent regulations on building projects. JLL’s Green predicts as more data centers come on line, hyperscalers will gobble up any new capacity.

He says cross-border data-sharing arrangements will drive new demand in Singapore and Hong Kong. In Singapore’s case, the government has banged out some data-sharing rules with Johor Bahru in Malaysia, next door. But the biggest shock will be the Greater Bay Area. Green notes that Hong Kong and mainland authorities have agreed on a mechanism to share privacy data.

“Without this, there could not be a GBA,” Green said. “This is a gamechanger.” This will make it easier for mainland companies use Hong Kong as a springboard to globalize their businesses, in whatever industry.

Now the data-center industry will build out a vast capacity to meet GBA and AI-related demand. The fintech industry will be one beneficiary (and a significant customer). But will these new sites be designed to maximize electricity efficiency and to decarbonize? The track record so far is not encouraging, and the up-front costs of going green are relatively high. The data center industry, increasingly built for genAI, could hypercharge climate disaster as much as it does the digital economy.

DigFin direct!

Register to receive DigFin's newsletter

  • Hauptseite
  • Grocery Gourmet Food
  • Fintech’s data centers face a GenAI “energy paradox”