For two decades, data centres were relatively safe, lucrative real estate investments. They attracted investors because they served as digital infrastructure — the essential plumbing that keeps the internet running.
The growth of artificial intelligence (AI) and its increased demands on computing power has meant data centre builders are in a race to build out sufficient supply.
But in the past two weeks, investors have started to have genuine concerns about the long-term viability of these businesses. On Wednesday, shares in NextDC, one of Australia’s largest data centre builders, collapsed by 40% after it said delays at its flagship Melbourne site meant its revenue forecasts would fall short of forecasts.
These delays have come about after the local community raised significant concerns about the impact of NextDC’s M3 data centre expansion. Electricity distributor Jemena paused the construction of the giant concrete power poles that had been snaking their way along Footscray streets, after local councils raised concerns about the impact on residential neighbourhoods.
These 22-metre power poles are crucial to NextDC’s operations in Footscray. Data centre operators make their profits from selling access to the massive amounts of power that AI chips need. Without that power, these facilities cannot remain viable.
NextDC is not alone in facing significant headwinds. Data centre company Firmus abruptly cancelled its multi-billion dollar listing on the Australian stock exchange, due to a lack of investor demand.
Similarly, London-based data centre developer Nscale is struggling to attract buyers for a stock market listing on Wall Street. Shares in CoreWeave, which claims to operate the world’s most powerful AI supercomputer, dropped 15% in one day last week on concerns about its financial strength.
These jitters reflect the harsh reality of AI: while the hype is strong, the business models are coming under increasing scrutiny.
Rising costs for massive investments
Hundreds of billions of dollars have been spent globally on training models and building the data centres to house the infrastructure that powers AI. The pipeline of investments is expected to reach some US$10 trillion (A$14.25 trillion) by 2032.
Until recently, much of the capital expenditure was funded internally by the AI companies or by raising money directly from investors; more recently, they have been borrowing vast sums in the corporate bond market.
And while those investing in AI are patient, they still expect to see significant financial returns.
Jeff J Mitchell/Getty Images
Sticky inflation and rising interest rates complicate this further. Data centre developers now face higher borrowing costs, plus higher construction costs. And rising interest rates can make other investments more attractive.
Analysts estimate that giants like OpenAI and Anthropic will need to increase revenues by 80% a year for the next six years to meet investors’ revenue expectations. Yet, indicators suggest that consumer and corporate spending on paid AI models is slowing down.
Racing for chips
To make things worse, the hardware that AI depends upon is driving real investor concern. Historically, data centres could operate for decades with minimal maintenance, guaranteeing rents and returns.
But AI chips have a shelf life of only a few years. Each new generation requires upgrades to power delivery and cooling, trapping data centres on a treadmill of multibillion-dollar expenses to pay for these upgrades.
Even if AI revenues stabilise enough to support these expenses, there’s no guarantee that data centre companies will be able to access the chips they need to fill data centres, and to support AI demand. Amazon recently admitted it is hoarding billions of AI chips, even though it doesn’t have any capacity for installing them. This massive over-investment is driven by anxieties around long-term supplies.
This gold rush for chips is why American chip manufacturer Nvidia is one of the few companies in the AI space to actually be making money. Nvidia also invests in the companies that buy its chips — its own customers — raising concerns about circular financing.
The social licence
Beyond the balance sheet, investors are also concerned about the loss of social licence for data centres amid local community protests. Community concerns around data centres are already high, driven by persistent noise (from energy transmission), visual impact, localised heating, and environmental concerns.
And these concerns will only grow if data centres start to impact the cost of living by pushing up electricity or water prices.
Forecasts suggest that New South Wales data centres could consume up to 25% of Sydney’s daily water demand within a decade.
In Victoria, transmission company AusNet is reviewing requests from data centres for more than 10 gigawatts of connections, which could exceed the state’s entire supply capacity.
In both cases, this level of demand will inevitably affect consumer prices.
When will communities benefit?
The AI age depends on the physical infrastructure of data centres. But the current path we are on demands ever-increasing capital from investors to fund a relentless upgrade cycle.
It demands huge volumes of power and water from local utilities. And it places a significant burden on the communities that host these data centres.
These factors mean the era of data centres being a safe stock market investment are over. And if investors become unwilling to fund these risks blindly, and communities become resistant to hosting these centres, this month’s stock market jitters may turn into a steeper market sell-off.
To shore up investor confidence, it’s important that investments in AI do not just focus on building better models, but also models that are more efficient — in terms of their costs, energy demands and spatial requirements.
And if they can be designed to support the electricity grid by adding, rather than displacing, renewables capacity, then communities may begin to feel the benefits of investments into AI.
Data centres have existed for decades. So why are they so controversial now?
