AI Data Centres Will Drive a 165% Power Demand: Explained

The AI boom is leading the global energy market to change, with data centres set to consume 165% more electricity by 2030 than they did in 2023, according to Goldman Sachs’ research.

The investment bank’s latest findings show that US data centre construction spending has tripled in just three years as tech giants rush to build the infrastructure needed for advanced AI systems. 

Despite new facilities opening regularly, occupancy rates at third-party data centres remain close to record highs across most US markets. 

James Schneider, Goldman Sachs’s Senior Equity Research Analyst covering digital infrastructure, says: “Over the next five to six years, we forecast substantial demand growth in the global data centre market.”

The scale of the transformation becomes clear when examining current power consumption patterns. 

Today’s global data centre market uses around 55GW of electricity, with more than half going to cloud computing services. 

Meanwhile, traditional business applications like email and file storage account for roughly a third, while AI workloads are just 14% of the total.

But by 2027, Goldman Sachs expects total consumption to jump to 84GW. 

How the power density revolution is changing the infrastructure game

Goldman Sachs believes the composition will change dramatically, with AI claiming more than a quarter of all power usage while cloud computing’s share drops to half and traditional workloads shrink to less than a quarter.

Where a traditional cloud server rack might draw a certain amount of power, its AI equivalent consumes ten times as much electricity. 

This intensity is only increasing – with power density across facilities expected to rise from 162 kilowatts per square foot today to 176 kilowatts by 2027.

“Data centre supply – specifically the rate at which incremental supply is built – has been constrained over the past 18 months,” James says – stemming from both the higher power requirements and the increased density of modern AI infrastructure.

Or in other words, a single question posed to ChatGPT, OpenAI’s conversational AI system, requires 2.9 watt-hours of electricity – nearly ten times what a Google search consumes, according to International Energy Agency data. 

Multiply that across millions of daily interactions – and the power implications become substantial.

Goldman Sachs models three scenarios for growth through 2027. 

The base case sees demand growing 50% to 92GW. Should AI adoption prove slower than expected, that rate could moderate to 14%. 

But if demand exceeds projections – perhaps through even more power-hungry chips or faster AI uptake – growth could accelerate to 20% annually.

The pattern mirrors earlier technology cycles, though the speed and scale feel unprecedented. 

Between 2015 and 2019, data centre workloads nearly tripled while power consumption stayed roughly flat at 200 terawatt-hours per year. 

Efficiency improvements offset the increased activity. Since 2020, however, those gains have slowed just as AI demand has accelerated, creating the current surge.

European market facing 170GW pipeline: explained

Goldman Sachs estimates Europe faces a data centre pipeline of roughly 170GW – equivalent to a third of the continent’s entire current electricity consumption.

Alberto Gandolfi, Managing Director, Equity Research at Goldman Sachs, says: “Inflecting power demand is monumentally important, because it’s been declining for 15 years in Europe.”

By 2030, European data centres alone will need as much electricity as Portugal, Greece and the Netherlands currently consume combined.

The geography of this expansion follows predictable patterns. 

Countries with cheap, abundant renewable energy – particularly the Nordic nations, Spain and France – are attracting investment based on their power advantages. 

Meanwhile, financial and technology hubs like Germany, Britain and Ireland are competing through tax incentives and their existing business ecosystems.

In Asia Pacific, established centres around Beijing and Shanghai continue expanding. 

While the region has added the most data centre capacity over the past decade, North America now leads in planned future development.

The market structure itself is consolidating around hyperscale operators – the handful of companies that run cloud services at massive scale. 

Amazon Web Services (AWS), Microsoft Azure and Google Cloud Platform are the largest players, though the category includes other major cloud providers. 

Goldman Sachs expects these hyperscalers and wholesale operators to control 70% of global capacity by 2030, up from 60% today.

Why grid strain demands US$720bn investment

Meeting this demand will require enormous infrastructure investment. 

US utilities alone need to spend US$50bn on new generation capacity just for data centres, while global grid upgrades could cost US$720bn through 2030.

“Retrofitting existing facilities to support these massive jumps in power density is becoming complex and compromised. We will need new, purpose-built AI infrastructure to power the next generation,” says Frank Long, a Vice President at the Goldman Sachs Global Institute, where he focuses on AI.

The timing compounds the difficulty, as efficiency improvements that previously masked growing power needs have largely plateaued.

Furthermore, data centre occupancy rates illustrate the supply-demand imbalance. 

From 85% in 2023, occupancy is projected to peak above 95% in late 2026 before moderating as new facilities come online. 

Such tight capacity utilisation levels typically signal market stress and rising prices.

The energy mix supporting this expansion is evolving towards renewables, though baseload power remains essential. 

Goldman Sachs forecasts 40% of new capacity will come from renewable sources. 

The economics appear compelling – as onshore wind costs around US$25 per megawatt hour and solar US$26, compared with US$37 for combined cycle natural gas before carbon costs.

Yet renewables face practical constraints. 

Utility-scale solar operates an average of six hours daily, wind plants hit nine hours. 

Data centres require continuous operation, pushing operators towards hybrid solutions that combine renewables with battery storage and backup natural gas capacity.

Nuclear power gaining support from the technology sector

The reliability demands of AI infrastructure are reviving interest in nuclear power. 

US technology companies have signed contracts for more than 10GW of potential new nuclear capacity over the past year, with Goldman Sachs identifying three plants that could begin operations by 2030.

The case for nuclear rests on its ability to provide consistent baseload power without carbon emissions. 

Large-scale onsite nuclear generation costs approximately US$77 per megawatt hour when carbon pricing reaches US$100 per tonne – competitive with gas-fired alternatives once environmental costs are included.

Brian Singer from Goldman Sachs Research says how technology companies’ sustainability commitments align with their reliability needs, creating momentum for both refurbishing decommissioned plants and building new reactors. 

The political environment has changed too, with nuclear enjoying bipartisan support in the US while countries like Switzerland reconsider their phase-out plans.

Furthermore, recent developments add uncertainty to these projections. 

DeepSeek, a Chinese AI model that reportedly matches leading US systems while using fewer computational resources, has raised questions about future infrastructure needs. 

The model’s efficiency gains, if proven scalable, could moderate the projected surge in power demand.

James acknowledges the uncertainty around DeepSeek’s “training, infrastructure and ability to scale” while maintaining Goldman Sachs’s broader forecasts. 

The firm notes that cooling systems account for 35-40% of hyperscaler energy consumption regardless of the underlying AI technology.

“Longer term, we see potential for a significant reduction of data centre emissions intensity and potentially in absolute emissions, as more nuclear power comes online and AI computing shifts to using AI models as opposed to training them,” the analysts say.