As generative AI asks for more power, data centers seek more reliable, cleaner energy solutions

Some parts of the world are already facing issues in generating power and managing grid capacity in the face of growing electricity demand from AI data centers.³ Critical power to support data centers’ most important components—including graphics processing unit (GPU) and central processing unit (CPU) servers, storage systems, cooling, and networking switches—is expected to nearly double between 2023 and 2026 to reach 96 gigawatts (GW) globally by 2026; and AI operations alone could potentially consume over 40% of that power.⁴ Worldwide, AI data centers’ annual power consumption is expected to reach 90 terawatt-hours by 2026 (or roughly one-seventh of the predicted 681 TWh of all data centers globally), roughly a tenfold increase from 2022 levels.⁵ As such, gen AI investments are fueling demand for so much electricity that in the first quarter of 2024, global net additional power demand from AI data centers was roughly 2 GW, an increase of 25% from the fourth quarter of 2023 and more than three times the level from the first quarter of 2023.⁶ Meeting data center power demand can be challenging because data center facilities are often geographically concentrated (especially in the United States) and their need for 24/7 power can burden existing power infrastructure.⁷

Deloitte predicts that both the technology and electric power industries can and will jointly address these challenges and contain the energy impact of AI—more specifically, gen AI. Already, many big tech and cloud providers are investing in carbon-free energy sources and pushing for net-zero targets,⁸ demonstrating their commitment to sustainability.

Hyperscalers plan massive expansion of gen AI data centers to help support growing customer demand

The surge in electricity demand is largely due to hyperscalers’ plans to build out data center capacity, globally.⁹ As AI demand—specifically gen AI—is expected to grow, companies and countries are racing to build more data centers to meet that demand. Governments are also establishing sovereign AI capabilities to maintain tech leadership.¹⁰ The data center real estate build-out has reached record levels based on select major hyperscalers’ capital expenditure, which is trending at roughly US$200 billion as of 2024, and estimated to exceed US$220B by 2025.¹¹

Moreover, Deloitte’s State of Generative AI in the Enterprise survey noted that enterprises have been mostly piloting and experimenting until now.¹² But as they experiment with getting value from gen AI, respondents are seeing tangible results and so intend to quickly scale up beyond pilots and proofs of concept. If usage grows as the technology matures, hyperscalers’ and cloud providers’ capital expenditure will most likely remain high through 2025 and 2026.

Two broad areas drive most of the electricity consumption in a data center: computing power and server resources like server systems (roughly 40% data center power consumption) and cooling systems (consume 38% to 40% power). These two are the most energy-intensive components even in AI data centers, and they will continue to fuel data centers’ power consumption. Internal power conditioning systems consume another 8% to 10%, while network and communications equipment and storage systems use about 5% each, and lighting facilities usually use 1% to 2% of power (figure 2).¹³ With gen AI requiring massive amounts of power, data center providers—including the hyperscalers and data center operators—need to look at alternate energy sources, new forms of cooling, and more energy-efficient solutions when designing data centers. Several efforts are already underway.

Gen AI is contributing to increased electricity demand

Data centers’ energy consumption has been surging since 2023, thanks to exploding demand for AI.¹⁴ Deploying advanced AI systems requires vast numbers of chips and processing capacity, and training complex gen AI models can require thousands of GPUs.

Hyperscalers and large-scale data center operators that are supporting gen AI and high-performance computing environments require high-density infrastructure to support computing power. Historically, data centers relied mainly on CPUs, which ran at roughly 150 watts to 200 watts per chip.¹⁵ GPUs for AI ran at 400 watts until 2022, while 2023 state-of-the-art GPUs for gen AI run at 700 watts, and 2024 next-generation chips are expected to run at 1,200 watts.¹⁶ These chips (about eight of them) sit on blades placed inside of racks (10 blades per rack) in data centers, and are using more power and producing more heat per square meter of footprint than traditional data center designs from only a few years ago.¹⁷ As of early 2024, data centers typically supported rack power requirements of 20 kW or higher. But the average power density is anticipated to increase from 36 kW per server rack in 2023 to 50 kW per rack by 2027.¹⁸

Total AI computing capacity, measured in floating-point operations per second (FLOPS), has also been increasing exponentially since the advent of gen AI. It’s grown 50% to 60% quarter over quarter globally since the first quarter of 2023 and will likely grow at that pace through the first quarter of 2025.¹⁹ But data centers don’t only measure capacity in FLOPS, they also measure megawatt hours (MWh) and TWh.

Gen AI’s multibillion parameter LLMs and the multibillion watts they consume

Gen AI large language models (LLMs) are becoming more sophisticated, incorporating more parameters (variables that enable AI learning and prediction) over time. From the 100 to 200 billion parameter models that were released initially during 2021 to 2022, current advanced LLMs (as of mid-2024) have scaled up to nearly two trillion parameters, which can interpret and decode complex images.²⁰ And there’s competition to release LLMs with 10 trillion parameters. More parameters add to data processing and computing power needs, as the AI model must be trained and deployed. This can further accelerate demand for gen AI processors and accelerators, and in turn, electricity consumption.

Moreover, training LLMs is energy intensive. Independent research of select LLMs that were trained on more than 175 billion parameters of data sets demonstrated that they consumed anywhere between 324 MWh and 1,287 MWh of electricity for each training run … and models are often retrained.²¹

On average, a gen AI–based prompt request consumes 10 to 100 times more electricity than a typical internet search query.²² Deloitte predicts that if only 5% of daily internet searches, globally, use gen AI– based prompt requests, it would require approximately 20,000 servers (with eight specialized GPU cores in each of the servers) that consume 6.5 kW on an average per server to fulfill the prompt requests, amounting to an average daily electricity consumption of 3.12 GWh and annual consumption of 1.14 TWh²³—which is equivalent to annual electricity consumed by approximately 108,450 US households.²⁴

Data center demand could present challenges and opportunities for power sector transition

The electric power sector was already planning for rising demand: Many in the industry predicted as much as a tripling of electricity consumption by 2050 in some countries.²⁵ But that trajectory has recently accelerated in some areas due to burgeoning data center demand. Previous forecasts in many countries have projected rising power demand due to electrification as well as increasing data center consumption and overall economic growth. But recent sharp spikes in data center demand, which could be just the tip of the iceberg, reveal the growing magnitude of the challenge.²⁶

This comes against the backdrop of a multi-decade power industry transformation, as electric companies build, upgrade, and decarbonize electric grid infrastructure, and digitalize systems and assets. In many areas, electric companies are also hardening assets against increasingly severe weather and climate events and protecting networks from rising cybersecurity threats.²⁷ Electric power grids in some countries are struggling to keep up with demand, especially for low- or zero-carbon electricity. In the United States, data centers are anticipated to represent 6% (or 260 TWh) of total electricity consumption in 2026.²⁸ The United Kingdom may witness a sixfold growth in electricity demand within a period of just 10 years, largely due to AI.²⁹ In China, data centers—including the ones that power AI—will likely make up 6% of the country’s total electricity demand by 2026.³⁰ Data centers could also add to China’s pollution problem, since the country’s power generation is dominated by coal, which accounted for 61% of its energy use and 79% of its carbon dioxide emissions in 2021.³¹

Some countries that are facing the rising demand for electricity from data centers are responding with regulations. For instance, in Ireland, existing data centers consume a fifth of the country’s total electricity consumption and this is only expected to grow further as AI-driven data centers spring up more; households are even lowering their power consumption.³² Temporarily, Ireland halted the construction of new data centers connected to the grid, but has since reversed that position.³³ Like Ireland, even the city of Amsterdam halted new data center construction to support sustainable urban development.³⁴ Singapore announced new sustainability standards for data centers that require operators to gradually increase the overall operating temperatures of their facilities to 26°C or higher. Higher operating temperatures reduce the demand for cooling and lower power consumption, but at the cost of shortening the lifespan of the chips.³⁵

The urgency and geographic concentration of data center demand—and the requirement for 24/7 carbon-free energy—can further complicate the challenge for tech companies and electricity providers, in addition to new demand from electrification, manufacturing, and other sources. The largest data center market globally is in northern Virginia,³⁶ and the local utility, Dominion Energy, expects power demand to grow by about 85% over the next 15 years, with data center demand quadrupling.³⁷ The round-the-clock, carbon-free electricity that many tech companies seek can be hard to come by, especially in the short term. Electricity providers are exploring multiple avenues to help meet demand while maintaining reliability and affordability. In addition to new renewable energy and battery storage, many electricity providers have also announced plans to build natural gas–fired power plants, which are not carbon-free.³⁸ This could potentially make it more challenging to meet utility, state, and even national decarbonization targets.³⁹

Despite being poised to consume massive amounts of clean energy, AI could also potentially help hasten the clean energy transition: Some utilities are already using AI to enable electric grids to operate more cheaply, efficiently, and reliably through improved weather and load forecasting, enhanced grid management and renewable asset performance, faster storm recovery, better wildfire risk assessment, and more.⁴⁰

Data center cooling is water-intensive

Next-generation CPUs and GPUs have higher thermal density properties than their predecessors. At the same time, some server vendors are packing more and more power-hungry chips into each rack in an endeavor to cater to the growing demand for high-performance computing and AI applications. But denser racks will demand more water, especially to cool the gen AI chips. AI data centers’ freshwater demand could be as much as 1.7 trillion gallons (at the higher end) by 2027.⁴¹ A hyperscale data center that intends to manage excess heat with air-based cooling and evaporated drinking water would require over 50 million gallons of water every year (or roughly what it takes to make 14,700 smartphones).⁴² This water cannot be returned to the aquifer, reservoir, or water supply where it came from.⁴³

Air-based cooling alone uses up to 40% of a typical data center’s electricity consumption. Therefore, data centers are looking at alternatives to traditional air-based cooling methods, mainly into liquid cooling, as its higher thermal transfer properties could help cool high-density server racks and enable them to reduce power usage by as much as 90% when compared with air-based methods.⁴⁴ As liquid cooling directly delivers cooling to server racks, it can support dense power racks on the order of 50 kW to 100 KW or more.⁴⁵ Moreover, it may help eliminate the need for chillers, which were traditionally used for producing the cooling water.

However, despite liquid cooling technology’s promise to help save energy across the data center stack,⁴⁶ it’s still in its early days and is yet to be widely adopted or integrated into AI data centers, globally.⁴⁷ Moreover, water is a finite resource, and therefore its cost and availability will likely affect future decisions about its usage. 

The tech industry is moving toward more sustainable solutions and carbon-free sources

To help expedite the move toward using carbon-free sources to power AI data centers, tech industry majors continue to be aggressive in their pursuit of renewable energy by way of power purchase agreements, or long-term contracts with renewable energy providers.⁴⁸ These deals have helped bankroll renewable energy projects by enabling them to secure financing. In some cases, technology companies are working with electricity providers and innovators to help test and scale promising energy technologies, including advanced geothermal, advanced wind and solar technologies, hydropower, and even underwater data centers.  

In some areas, local grid constraints and long interconnection times for new renewable and battery storage facilities are causing delays in connecting these resources to the electric grid.⁴⁹ These delays, which can be as long as five years in the United States, are often due to high demand and insufficient transmission infrastructure. As a result, tech companies are increasingly pursuing onsite, sometimes off-grid, energy solutions.⁵⁰ Additionally, they are investing in new technologies such as long-duration energy storage and small modular nuclear reactors to help address these challenges. In some cases, tech companies and utilities are planning to coordinate to bring innovative clean energy technologies to scale, which could eventually benefit other organizations and help decarbonize the electric grid faster.⁵¹ Many of these research and development programs, pilots and other clean energy investments may take years before reaping benefits, demonstrating return on investment, and becoming commercially viable.⁵² For example, small, modular nuclear reactors are still in early development stages, and may not be a near-term zero-carbon solution.⁵³

The technology sector consistently dominates US corporate renewable procurement and accounted for more than 68% of the nearly 200 deals of associated contracted capacity tracked over the 12 months prior to Feb. 28, 2024.⁵⁴ Similarly, hyperscalers and data center operators in India are increasingly using solar to power their data centers in the region.⁵⁵ Without these purchase commitments, many renewable energy projects would not be built.⁵⁶

As such, the tech industry’s role in bankrolling clean energy technologies to help bring them to scale will continue to be valuable. In some cases, they’re working directly with innovators and renewable energy producers, and in other cases, they’re partnering with utilities.⁵⁷ Importantly, the way tech companies inject capital to help advance the clean energy transition is critical, as neither the innovators nor the power industry would typically have the level of financial resources that the tech industry possesses.

Bottom line

What should the broader tech industry, hyperscalers, data center operators, utilities, and regulators do to help meet gen AI demand sustainably? Several considerations for hyperscalers and the broader tech industry are more or less in line with what we presented in Deloitte Global’s 2021 prediction on cloud migration.⁵⁸ Though demand drivers may have changed and the pace of change has accelerated, the industry is working to achieve balance between sustainability and expedited time to market, while keeping data centers rising energy demands under control and finding more sustainable ways to power AI—specifically gen AI.

1. Make gen AI chips more energy-efficient: Already, a new generation of AI chips can perform AI training in 90 days, consuming 8.6 GWh. This would be less than one-tenth the energy that the previous generation of chips takes to do the same function on the same data.⁵⁹ Chip companies should continue to work with the broader semiconductor ecosystem to help intensify focus on improving FLOPS/watt performance, such that, future chips can train AIs several times larger than the largest AI systems currently available while using less electricity.

2. Optimize gen AI uses and shift processing to edge devices: This includes assessing whether it’s energy-efficient to do training and inference in the data center, or on the edge device, and accordingly rebalance data center equipment needs. Edge not only can support applications where response times are critical, but also for those use cases where sensitive data is involved, or privacy needs are high. It also helps save network and server bandwidth, rerouting gen AI workloads to local and near-location or co-location devices, while only transmitting select AI workloads to data centers.⁶⁰

3. Implement changes in gen AI algorithms and rightsize AI workloads: Do we need to keep racing to build bigger and bigger foundational models (for example, more than a trillion parameter models), or are smaller models sufficient, while being more sustainable? Already, startups are developing on-device multimodal AI models, which do not require energy-intensive computations in the cloud.⁶¹ Customers should fine-tune and adjust the size of their AI workloads and go for targeted gen AI models (including preexisting models and training only when needed) based on real business needs, which can minimize energy use. Additionally, depending on specific needs with AI inferencing (for example, doing inference in real time and when latency is critical), CPUs can be more advantageous and efficient.⁶²

4. Form strategic partnerships to serve local and cluster-level AI data center needs: For several small to midsize organizations (including universities) that may find it hard to tap into gen AI data center capacity, those organizations should work with specialized data center operators and cloud service providers that focus on delivering high-performance computing solutions for smaller high-performance computing GPU-cluster co-locations.⁶³ A corollary: Data centers can then actively track usage and availability for potential opportunities and demand pockets to help deliver near-term co-location services.

5. Collaborate with other stakeholders and sectors to make an overall positive environmental impact: The various ecosystem players—including hyperscalers, their customers, third-party data center operators and co-location service providers, electricity providers, the local regulators and municipalities, and the real estate firms—should have ongoing conversations around what’s feasible and viable for the business, environment, and society.⁶⁴ That collaboration should encompass multiple aspects including determining potential strategic co-location needs (where a data center company rents computing and server resources to one or more companies), assessing cooling needs such as adequate temperatures in liquid cooling systems, identifying solutions to manage heat and wastewater, and figuring out recycling needs. For example, in Europe, data center operators are directing waste heat to warm swimming pools in the vicinity.⁶⁵ Electricity providers should consider working more closely with the tech industry to understand how to meet data center energy demand, while identifying ways the tech companies could potentially help fund and scale new energy technologies, which is a vital step to bring more clean energy to the grid.

The holistic efforts of hyperscalers and electricity providers to help increase the use of carbon-free sources to power data centers—including the ones being built exclusively for gen AI—may bear fruit in the longer term.