Generative AI’s demand for computing power will be limited by the inability of power utilities to expand their generation capacity fast enough and data centre administrators must include power supply limitations into expansion, pricing models and workload planning says Bob Johnson at Gartner.
The explosive growth of new hyperscale data centres to implement Generative AI is creating an insatiable demand for power that will exceed the power utilities’ ability to provide it. New, bigger data centres are being planned to handle the huge amounts of data needed to train and implement the continuously expanding large language models that underpin popular Generative AI applications. Data centre managers are realizing that there will not be enough power available to implement their most ambitious plans.
This problem is exacerbated by regional concentrations of data centres, which can consume a major percentage of the local power utility’s total production. Such concentrations benefit from well-established communication infrastructure, but can also threaten to overwhelm the power utility’s future capacity with expansion plans.
The situation is serious enough that some locations, such as Ireland and Singapore, are either severely limiting or completely stopping data centre growth out of power concerns.
Cost of power
The inevitable result of impending power shortages is an increase in the price of power, which will also increase the costs of operating Generative AI’s large language models. Data centre operators can expect major cost increases as the largest data centre operators compete financially for scarce power resources.
Data centres that demand huge amounts of power can be built far faster than power utilities can expand their capacity. Delivering increased power to data centre locations often requires new transmission lines from existing generation facilities, which can take years for permits, but can even necessitate building new power plants, which can take decades.
Short-term solutions to provide more power often conflict with zero-carbon sustainability goals. Data centres require 24×7 power availability, which renewable power such as wind or solar cannot provide without some form of alternative supply during periods when they are not generating power. Truly 24×7 reliable power can only be generated by either hydroelectric power, fossil fuel or nuclear power plants.
Hydropower is limited by appropriate sites, and fossil fuel plants that had been scheduled to be retired are being kept operating in response to increased data centre power needs.
The growth in number and computational power of Generative AI data centres, along with the infrastructure that supports them, requires ever increasing amounts of electrical power. Demand is quickly approaching the limits of the global electrical power industry’s ability to provide sufficient generating capacity.
The total estimated power demand for the collection, analysis, computation, storage and communication for global information and communication technology, is growing significantly faster than the total electrical energy production. Total ICT electricity demands will exceed 9% of total global energy production by 2030, up from less than 3% today.
To satisfy this increase in demand, ICT applications would have to take considerable amounts of electricity from other areas, such as residential, commercial or industrial users. This is an untenable situation.
The total demand for ICT power is projected to grow at a 25% compound annual growth rate, CAGR through 2030, while global power generation will only increase at a 3% CAGR. This means that while total ICT power demand accounts for less than 3% of the total power generation in 2024, it will account for over 9% by 3030.
In 2021, the additional demand for ICT accounted for a manageable 4% of the additional generation capacity globally. By 2023, driven by the increase in hyperscale data centres, incremental ICT demand was 20% of total new generation, and this is projected to exceed 70% by 2030.
While these numbers may seem alarming on a global basis, they can be amplified in regions that have a large concentration of data centres.
Demand and supply imbalance
The result is that if current trends in data centre construction continue, a shortage in available power from traditional grid suppliers will rapidly develop. While this shortage will become acute by the end of the decade, regional shortfalls are likely to occur on a significant level within a couple of years.
In past years, ICT power consumption could grow much faster than electricity supply because it was a relatively small portion of the total. However, as ICT exceeds 5% of the total supply, it becomes a significant factor in electricity consumption. ICT will begin to compete with other major users like residential power, about 40%, commercial power, about 35% or industrial power, about 25% for access to power generation capacity.
The inevitable result of the impending electricity supply-demand imbalance is that, as demand for electricity begins to outstrip supply, electricity will get more expensive. Consequently, marginal applications will drop out of the market, and only those that can afford higher-priced electricity will remain.
In addition, as power scarcities begin to become apparent, regional planning authorities will limit approvals for new data centres and other ICT infrastructure that would strain existing power grids. The result is that data compute, storage and communication capacity becomes supply-constrained and, therefore, expensive.
However, an added complication is that, if electricity gets too expensive for average consumers, politicians may force the allocation of electricity to consumer applications like powering the home and transportation. This forced allocation will reduce the amount of electricity available to other applications, like ICT.
The impact of impending power scarcities will be greater in areas with large concentrations of data centres. In some areas, wholesale power prices — the price for power direct from the generating plant into the distribution grid — are already reacting to potential shortages.
Major users of power are working with major power producers to secure long-term guaranteed sources of power independent of other grid demands. As an example, Amazon Web Services, AWS recently acquired a 1,200-acre data centre campus from Talen Energy. The facility, which can utilize up to 960MW of power, is adjacent to the Susquehanna nuclear power station.
A key aspect of this deal is that the data centre gets guaranteed power directly from the power station, bypassing the traditional grid. Because the nuclear power facility also supplies power to the general grid, the deal is being challenged by other users who claim that it will raise prices for other users.
Green energy
Data centres require 24×7 reliable power and cannot afford to be affected by unscheduled brown-outs or blackouts or interruptions of their baseload power. Currently, most data centres get their power from regional power grids that have a variety of power sources and can provide the type of reliable power needed. However, the surging demand for power for new data centres is forcing public utility power suppliers to increase their production by any means possible.
In some cases, this means keeping fossil fuel plants that had been scheduled for retirement operating beyond their scheduled shutdown. Other examples include running additional long distance power lines to existing fossil fuel plants or keeping such plants already connected to the grid running longer. While many of these strategies are successful in providing needed power to new data centres, an undesirable side effect is an increase in the carbon footprint.
While renewable power, wind and solar is capturing an increasing percentage of total generating capacity, these sources are not suitable as a primary source of data centre power without some form of reliable, 24×7 backup. The difficulty with renewable power is that both wind and solar depend on factors they cannot control, wind and sunshine and can only be depended upon to produce a fraction of their total rated power over an extended period of time.
The capacity utilization factor, CUF is a number derived by taking the actual power produced by a type of generation over a period of time, like a year, and dividing it by the maximum rated power a given installation can produce. Wind and solar have relatively low CUFs — around 25% for solar and 35-40% for wind.
This means that while renewables are an essential contributor of power to a wide ranging grid where power peaks during the day, they cannot be depended on to provide the type of power needed for data centre operation. This is also one of the primary reasons some data centres are instilling dedicated microgrids or hybrid solutions to allow for on-site generation during periods when renewable power is not available.
Utility-scale batteries are emerging as a viable short-term energy storage medium to level grid load, especially in periods of high demand, like for air conditioning during heat waves. However, batteries are expensive, and while large scale installations can provide hyperscale levels of power, they can only do it for short periods of time.
The reality is that in the short term, increased data centre use will lead to increased CO2 emissions to generate the needed power. This, in turn, will make it more difficult for data centre operators and their customers to meet aggressive sustainability goals relating to CO2 emissions. Longer term, new technologies for improved battery storage, like sodium ion batteries, or clean power, like small nuclear reactors, will become available and help achieve sustainability goals.
Source: Power Shortages Will Restrict Generative AI Growth and Implementation by Gartner
Medium to long term plans for data centre administrators supporting Generative AI workloads
- Evaluate the impact of potential power shortages on products and services by developing scenarios that reflect limitations on available power.
- Include significant cost increases for data centre services when developing plans for new products and services.
- Look for alternative approaches that require less power while achieving the same market impact.
- Reevaluate sustainability goals for CO2 production in light of anticipated increases in the use of fossil-fuel-based power generation.
- Determine the risks and impact potential power shortages will have on all products and services, current and planned, by evaluating the likelihood that they will require new data centre construction to be successful.
- Develop workaround solutions for power shortages, outages by identifying major risk points and potential alternatives.
- Diversify to geographic regions with sufficient power to ensure that products and services will not be affected by power shortages.
- Negotiate long-term contracts for data centre services at reasonable rates for power.
- Conduct a cost-benefit analysis of planned expansions of products and services in light of projected data centre costs.
- Consider moving some products and services to regional data centres in areas with more predictable power costs.
- Reevaluate sustainability goals relating to CO2 emissions in light of future data centre requirements.
- Work with data centre providers and power utilities to evaluate the emissions potential for data centre alternatives.
- Develop Generative AI applications with a focus on using a minimum amount of computing power.
- Look at the power, and emissions implications of options such as edge computing and smaller language models.
- Support initiatives to promote development and implementation of green power alternatives such as nuclear, green hydrogen, fusion, geothermal.
- Reevaluate sustainability goals relating to CO2 emissions in light of future data centre requirements and data centre power sources.
Near-term plans for data centre administrators supporting Generative AI workloads
- By 2026, half of new data centre construction will have dedicated on-site power generation capabilities to supplant grid power if necessary.
- Power shortages will limit growth of new data centres for Generative AI and other uses starting in 2026.
- As demand for electricity exceeds supply, electricity prices will increase until a new supply-demand balance is achieved.
- Sustainability goals will be negatively affected as new fossil fuel plants will be required for reliable electricity supply.
- The number of new data centres, and growth of large language models, will be governed by availability of power to run them.
- The potential for power shortages is higher in regions with significant concentrations of data centres.
- More attractive locations for new data centres may be found in regions with few data centres.
- These areas are more likely to have the potential for surplus power available.
- Some data centre providers are working with power suppliers to lock in guaranteed supplies on long-term purchase agreements.
- Expect the cost of power to operate data centres to increase significantly as data centre operators use economic leverage.
- Increased power costs will be passed on to AI, Generative AI product and service providers as well.
- Regional power costs are likely to fluctuate significantly in areas with large concentrations of data centres competing for available power.
- Short-term power shortages are likely to continue for years.
- New power transmission, distribution and generation capacity takes years to come on line and will not alleviate current problems.
- Sustainability targets relating to reducing CO2 emissions for the near term, three to five years will likely be missed if plans call for rapidly increasing data centre use, especially for compute-heavy Generative AI applications.
- Longer-term green alternative ways to generate power will become available, but these will take years to come to fruition.
- Adopting Generative AI applications that require less compute power will help meet sustainability goals.
