The use of artificial intelligence is growing, leading to increased energy demands in data centers. Experts warn that the electricity consumption of entire countries could be affected.
According to Ralf Herbrich, the director of the Hasson Plattner Institute (HPI) in Potsdam and head of the artificial intelligence and sustainability department, the energy consumption of AI tools is substantial and on the rise. The process of managing a single AI model requires a significant amount of energy due to complex prediction calculations.
Alex de Vries, a data scientist from Amsterdam, has compared the energy consumption of AI-powered search engines to that of entire countries. This issue is becoming increasingly important for climate protection. Efforts are being made by scientists and internet companies to reduce the ecological impact of AI.
Ralf Herbrich mentioned that data centers currently account for four to five percent of global energy consumption, and this figure rises to eight percent when including the use of digital technologies like laptops and smartphones. It is estimated that this consumption could increase to 30 percent in the coming years.
To train an AI model, hundreds of graphics cards’ processors, each consuming around 1,000 watts, run for several weeks. Herbrich compared this to an oven, stating that 1,000 watts is as much as an oven consumes.
The topic of artificial intelligence is currently a dominant factor in public discussions about technology. It has gained considerable attention, especially due to the text robot ChatGPT from the Californian startup OpenAI. AI applications are becoming more widespread, including safety technology in cars and efficient heating systems, as well as various applications in healthcare and other industries.
Efforts are being made to reduce the energy consumption of AI technology while maintaining the accuracy of predictions. It will take several years to develop solutions, according to Herbrich from the Hasso Plattner Institute. Technology companies are also actively researching energy-efficient AI.
Researcher de Vries estimates that if every Google search utilized AI, it would require around 29.2 terawatt hours of electricity per year, equivalent to Ireland’s annual electricity consumption. However, this is viewed as an extreme scenario that is unlikely to occur in the near term.
Google states that the energy required to operate their AI technology is increasing at a slower pace than many had predicted. They have employed proven methods to significantly reduce the energy consumption for training AI models. Additionally, Google uses AI for climate protection, such as for “fuel-efficient route planning” on Google Maps and predicting river flooding.
In various industries, the rising demand for energy, mainly from the construction and operation of data centers used for training and running AI models, is contributing to global greenhouse gas (GHG) emissions. Microsoft, which has invested in OpenAI, the maker of ChatGPT, and has placed generative AI tools at the core of its product offering, recently declared that its CO2 emissions had increased by almost 30% since 2020 due to the expansion of data centers. Google’s GHG emissions in 2023 were nearly 50% higher than in 2019, largely because of the energy demand related to data centers.
While AI tools pledge to aid in the energy transition, they also necessitate substantial computing power. The energy consumption of AI currently represents only a small part of the technology sector’s power usage, estimated to be approximately 2-3% of total global emissions. It is probable that this will change as more companies, governments, and organizations utilize AI to drive efficiency and productivity. As shown by this chart, data centers are already significant drivers of electricity demand growth in many regions.
AI requires significant computing power, and generative AI systems may already consume about 33 times more energy to complete a task than task-specific software. With the increasing adoption and advancement of these systems, the training and operation of the models will lead to a substantial escalation in the required number of global data centers and associated energy usage. Consequently, this will exert additional pressure on already overburdened electrical grids.
Notably, training generative AI is exceptionally energy-intensive and consumes a much greater amount of electricity compared to traditional data center activities. As an AI researcher articulated, “When you deploy AI models, you have to have them always on. ChatGPT is never off.” The growing sophistication of a large language model, like the one on which ChatGPT is constructed, serves as evidence of this escalating energy demand.
Training a model such as Generative Pre-trained Transformer 3 (GPT-3) is believed to consume just under 1,300 megawatt hours (MWh) of electricity, roughly equivalent to the annual power consumption of 130 homes in the US. Meanwhile, training the more advanced GPT-4 is estimated to have utilized 50 times more electricity.
Overall, the computational power essential for supporting AI’s growth is doubling approximately every 100 days. Society therefore contends with challenging questions, pondering whether the economic and societal benefits of AI outweigh its environmental cost. Specifically, the inquiry arises as to whether the benefits of AI for the energy transition outweigh its heightened energy consumption.
The quest for the optimal balance between challenges and opportunities is crucial for obtaining the answers we seek. Reports forecast that AI has the potential to mitigate 5-10% of global GHG emissions by 2030. Thus, what needs to happen to strike the right balance?
Regulators, including the European Parliament, are commencing efforts to establish requirements for systems to be designed with the ability to record their energy consumption. Furthermore, technological advancements could mitigate AI’s energy demand, with more advanced hardware and processing power anticipated to enhance the efficiency of AI workloads.
Researchers are crafting specialized hardware, such as new accelerators, as well as exploring new technologies like 3D chips that offer significantly improved performance, and novel chip cooling techniques. Nvidia, a computer chip manufacturer, asserts that its new ‘superchip’ can achieve a 30 times improvement in performance when operating generative AI services while consuming 25 times less energy.
Concurrently, data centers are becoming more efficient, with ongoing exploration into new cooling technologies and sites capable of executing more computations during periods of cheaper, more available, and sustainable power to further advance this efficiency. Alongside this, reducing overall data usage, including addressing the phenomenon of dark data — data generated and stored but then never used again — is crucial. Additionally, being more selective about how and where AI is used, for instance, by employing smaller language models, which are less resource-intensive, for specific tasks will also contribute. Striking a better balance between performance, costs, and the carbon footprint of AI workloads will be fundamental.
What about AI’s impact on the electrical grid? AI is not the sole factor applying pressure to the grid. Increasing energy needs due to growing populations, as well as trends toward electrification, are creating heightened demand that could result in a slower decarbonization of the grid.
Nonetheless, a clean, modern, and decarbonized grid will be imperative in the broader shift to a net-zero emissions economy. Data center operators are exploring alternative power options, such as nuclear technologies for powering sites, or storage technologies like hydrogen. Additionally, companies are investing in emerging technologies, such as carbon removal, to extract CO2 from the air and store it securely.
AI can help overcome obstacles to integrating the necessary large amounts of renewable energy into existing grids.
The fluctuation in renewable energy generation often leads to excess production during peak times and shortages during lulls, causing inefficient energy usage and unstable power grids. By analyzing large sets of data, ranging from weather patterns to energy consumption trends, AI can accurately predict energy production. This could facilitate scheduling tasks and shifting loads to ensure that data centers use energy when renewable energy sources are available, thus ensuring stable grid operations, efficiency, and continuous clean power. AI is also aiding in improving the energy efficiency of other industries that produce large amounts of carbon, from analyzing buildings to anticipate energy usage and optimize heating and cooling system performance to enhancing manufacturing efficiency with predictive maintenance. In agriculture, sensors and satellite imagery are being used to forecast crop yields and manage resources.
Effectively managing the energy consumption and emissions of AI while maximizing its societal benefits involves addressing multiple interconnected challenges and requires input from various stakeholders.
The World Economic Forum’s Artificial Intelligence Governance Alliance is examining how AI can be utilized in different industries and its impact on innovation, sustainability, and growth.
As part of this effort, the Forum’s Centre for Energy and Materials and Centre for the Fourth Industrial Revolution are launching a specific workstream to explore the energy consumption of AI systems and how AI can facilitate the transition to clean energy.
In an era where the rapid advancements in Artificial Intelligence (AI) captivate society, the environmental impact of these advancements is often disregarded. The significant ecological consequences of AI demand attention and action.
For AI to realize its potential for transformation, offering unprecedented levels of productivity and enhancing societal well-being, it must develop sustainably.
At the core of this challenge is the significant energy demand of the AI ecosystem, encompassing everything from hardware to training procedures and operational methods.
Notably, the computational power required to sustain the rise of AI is doubling approximately every 100 days. To achieve a tenfold improvement in AI model efficiency, the demand for computational power could increase by up to 10,000 times. The energy required to perform AI tasks is already increasing at an annual rate of between 26% and 36%. This means that by 2028, AI could be utilizing more power than the entire country of Iceland did in 2021.
The environmental impact of the AI lifecycle is significant during two key phases: the training phase and the inference phase. During the training phase, models learn and improve by processing large amounts of data. Once trained, they move into the inference phase, where they are used to solve real-world problems. Currently, the environmental impact is divided, with training accounting for about 20% and inference consuming the majority at 80%. As AI models gain traction across various sectors, the need for inference and its environmental impact will increase.
To align the rapid progress of AI with the imperative of environmental sustainability, a carefully planned strategy is crucial. This entails immediate and near-term actions while also establishing the groundwork for long-term sustainability.
Immediate Approach: Reducing AI’s energy consumption today
Research is emerging about the practical steps we can take now to align AI progress with sustainability. For instance, capping power usage during the training and inference phases of AI models provides a promising avenue for reducing AI energy consumption by 12% to 15%, with a marginal tradeoff in task completion time, as GPUs are expected to take around 3% longer.
Another impactful method is optimized scheduling for energy conservation. Tasking AI workloads to align with periods of lower energy demand — such as running shorter tasks overnight or planning larger projects for cooler months in regions where air conditioning is widely used — can also result in significant energy savings.
Finally, transitioning towards the use of shared data centers and cloud computing resources instead of individually setting up private infrastructure can concentrate computational tasks in collective infrastructures and reduce the energy consumption associated with AI operations. This can also lead to cost savings on equipment and potentially lower energy expenses, particularly when resources are strategically placed in areas with lower energy costs.
Near-Term Focus: Utilizing AI for the energy transition
Beyond immediate measures, the near-term focus should be on leveraging AI’s capabilities to promote sustainability. AI, when used effectively, can be a powerful tool in meeting the ambitious goal of tripling renewable energy capacity and doubling energy efficiency by the end of the decade, as established in last year’s United Nations Climate Change Conference (COP28).
AI supports climate and energy transition efforts in various ways. It assists in developing new materials for clean energy technologies and optimizing solar and wind farms. AI can also enhance energy storage capabilities, improve carbon capture processes, and refine climate and weather predictions for better energy planning, as well as stimulate innovative breakthroughs in green energy sources like nuclear fusion.
Strategically using AI to improve our renewable energy landscape offers the promise of not only making AI operations environmentally friendly, but also contributing to the creation of a more sustainable world for future generations.
In the long run, creating synergy between AI and emerging quantum technologies is a crucial approach to guiding AI toward sustainable development. Unlike traditional computing, where energy usage increases with greater computational demand, quantum computing shows a linear relationship between computational power and energy consumption. Furthermore, quantum technology has the potential to transform AI by making models more compact, improving their learning efficiency, and enhancing their overall functionality, all without the significant energy footprint that is currently a concern in the industry.
Realizing this potential requires a collective effort involving government support, industry investment, academic research, and public engagement. By combining these elements, it is conceivable to envision and establish a future where AI advances in harmony with the preservation of the planet’s health.
Standing at the intersection of technological innovation and environmental responsibility, the way forward is clear. It requires a collective effort to embrace and propel the integration of sustainability into the core of AI development. The future of our planet depends on this crucial alignment. Decisive and collaborative action is necessary.
Global spending on offshore energy infrastructure over the next decade is projected to exceed US$16 billion (£11.3bn), which includes laying an additional 2.5 million kilometers of global submarine cables by 2030.
The process of laying and securing these cables against ocean currents involves disturbing the seabed and depositing rocks and concrete “mattresses” to serve as a base for the cables. These procedures can have a significant impact on the marine ecosystem, which is home to numerous creatures.
The installation of offshore wind farms entails many high-impact procedures that are often carried out with little consideration for their effects on the delicately balanced ocean environment, which supports the food and livelihoods of over 3 billion people.
Human activities, including the construction of renewable offshore energy infrastructure, have impacted over 40% of the ocean’s surface, leading to dead ocean zones devoid of oxygen, harmful algae blooms, and a devastating loss of biodiversity.
If we continue on this trajectory, the anticipated green-tech revolution risks causing an unprecedented level of harm to the world’s oceans. The new generation of renewable energy producers needs to evaluate the long-term impact of their actions on the ocean environment to determine the true sustainability of their supply chains and practices.
As the UN commences its decade of Ocean Resilience this year, the role that autonomous technologies can play in supporting the marine environment is increasingly gaining recognition. Implementing sustainable technology necessitates instilling environmentally conscious practices within the renewable energy sector itself. This is where robotics can contribute.
Approximately 80% of the cost of maintaining offshore wind farms is allocated to sending personnel for inspections and repairs via helicopter, maintaining support vehicles such as boats, and constructing offshore renewable energy platforms to accommodate turbine workers. All of these activities contribute to carbon emissions, and they also pose risks to human safety.
However, a unified team of humans, robots, and AI working together could maintain this infrastructure with significantly less impact on the environment and better safety for humans. Such teams could involve humans working remotely with multi-robot teams of autonomous aerial and underwater vehicles, as well as with crawling or land-based robots.
Robotic technology can enable humans to interact with complex and vulnerable environments without causing harm. Robots equipped with non-contact sensing methods, such as radar and sonar, can interact with ocean infrastructure and its surrounding environment without causing any disruption or damage.
Even more advanced sensing technology, inspired by the communication signals used by dolphins, makes it possible to inspect structures such as subsea infrastructure and submarine cables in the ocean without harming the surrounding environment.
Using autonomous underwater vehicles (AUVs) that operate independently, we can gain a better understanding of how offshore energy structures, like underwater cables, interact with the environment, through the deployment of low-frequency sonar technology. This technology can also assist in preventing issues such as biofouling, where microorganisms, plants, algae, or small animals accumulate on the surfaces of cables.
Biofouling can cause a bio-fouled cable to become heavy, potentially distorting its outer protective layers and reducing its useful life span. AUVs have the capability to monitor and clean these cables safely.
Robotic assistance can also be extended to offshore energy infrastructure above the water. When wind turbine blades reach the end of their useful lives, they are often incinerated or disposed of in landfills. This practice contradicts the principles of the “circular economy,” which emphasizes waste prevention and the reuse of materials for sustainability. Instead, robots can be employed to repair, repurpose, or recycle deteriorating blades, thereby reducing unnecessary waste.
Advanced radar sensing technology mounted on drones enables us to detect defects in turbines as they start to develop. By utilizing robot assistants to stay updated on turbine maintenance, we can avoid the need for costly field support vessels to transport turbine inspectors offshore, which can amount to around £250,000 a day. This approach helps in saving time, money, and reducing risk.
In addition to cutting the financial and carbon cost of turbine maintenance, robots can also minimize the inherent risks to humans working in these unpredictable environments, while operating more harmoniously with the environment. By deploying resident robots for the inspection and maintenance of offshore renewable infrastructure, energy companies could initially decrease the number of people working in hazardous offshore roles. Over time, this could lead to autonomous operation, where human operators remain onshore and connect remotely to offshore robotics systems.
AI plays a significant role in the establishment of sustainable offshore energy systems. For instance, artificially intelligent programs can aid offshore energy companies in planning the safe disassembly and transportation of turbines back to shore. Upon arrival onshore, turbines can be taken to “smart” factories that utilize a combination of robotics and AI to identify which parts can be reused.
By collaborating in these efforts, we can develop a resilient, sustainable circular economy for the offshore renewable energy sector.
The latest IPCC report is clear: urgent action is needed to avoid severe long-term climate effects. Given that more than 80% of global energy still comes from fossil fuels, the energy sector must play a central role in addressing this issue.
Thankfully, the energy system is already undergoing a transformation: renewable energy production is rapidly expanding due to decreasing costs and growing investor interest. However, the scale and cost of decarbonizing the global energy system are still enormous, and time is running out.
Thus far, most of the efforts to transition the energy sector have focused on physical infrastructure: new low-carbon systems that will replace existing carbon-intensive ones. Comparatively little effort and investment have been directed toward another crucial tool for the transition: next-generation digital technologies, particularly artificial intelligence (AI). These powerful technologies can be adopted on a larger scale and at a faster pace than new physical solutions and can become a crucial enabler for the energy transition.
Three significant trends are propelling AI’s potential to expedite the energy transition:
1. Energy-intensive sectors like power, transportation, heavy industry, and buildings are at the outset of transformative decarbonization processes driven by increasing government and consumer demands for rapid CO2 emission reductions. The scale of these transitions is immense: BloombergNEF estimates that achieving net-zero emissions in the energy sector alone will necessitate between $92 trillion and $173 trillion of infrastructure investments by 2050. Even slight gains in flexibility, efficiency, or capacity in clean energy and low-carbon industry can result in trillions of value and savings.
2. As electricity powers more sectors and applications, the power sector is becoming the cornerstone of global energy supply. Scaling up the deployment of renewable energy to decarbonize the expanding power sector globally will result in a greater portion of power being supplied by intermittent sources (such as solar and wind), creating new demand for forecasting, coordination, and flexible consumption to ensure the safe and reliable operation of power grids.
3. The transition to low-carbon energy systems is fueling the rapid expansion of distributed power generation, distributed storage, and advanced demand-response capabilities, which need to be coordinated and integrated through more interconnected, transactional power grids.
Navigating these trends presents significant strategic and operational challenges to the energy system and energy-intensive industries. This is where AI comes in: by establishing an intelligent coordination layer across energy generation, transmission, and utilization, AI can assist energy-system stakeholders in identifying patterns and insights in data, learning from experience, enhancing system performance over time, and predicting and modeling potential outcomes of complex, multivariate scenarios.
AI is already demonstrating its value to the energy transition in various areas, driving verifiable enhancements in renewable energy forecasting, grid operations and optimization, coordination of distributed energy assets and demand-side management, and materials innovation and discovery.
While AI’s application in the energy sector has shown promise thus far, innovation and adoption are still limited. This presents a significant opportunity to expedite the transition toward the zero-emission, highly efficient, and interconnected energy system needed in the future.
AI holds far greater potential to expedite the global energy transition, but realizing this potential will only be achievable through greater AI innovation, adoption, and collaboration across the industry. This is why the World Economic Forum has published ‘Harnessing AI to Accelerate the Energy Transition,’ a new report aimed at defining and catalyzing the necessary actions.
The report, developed in collaboration with BloombergNEF and Dena, establishes nine ‘AI for the energy transition principles’ targeting the energy industry, technology developers, and policymakers. If implemented, these principles would hasten the adoption of AI solutions that support the energy transition by establishing a shared understanding of what is required to unlock AI’s potential and how to adopt AI in the energy sector in a safe and responsible manner.
The principles define the actions needed to unlock AI’s potential in the energy sector across three vital domains:
1. Governing the use of AI:
Standards – implement compatible software standards and interoperable interfaces.
Risk management – agree on a common approach to technology and education to manage the risks posed by AI.
Responsibility – ensure that AI ethics and responsible use are at the heart of AI development and deployment.
2. Designing AI that’s fit for purpose:
Automation – design generation equipment and grid operations for automation and increased autonomy of AI.
Sustainability – adopt the most energy-efficient infrastructure as well as best practices for sustainable computing to reduce the carbon footprint of AI.Design – focus AI development on usability and interoperability.
3. Facilitating the implementation of AI on a large scale:
Data – establishing standards for data, mechanisms for sharing data, and platforms to enhance the availability and quality of data.
Education – empowering consumers and the energy workforce with a human-centered approach to AI and investing in education to align with technological advancements and skill development.
Incentives – developing market designs and regulatory frameworks that enable AI use cases to capture the value they generate.
AI is not a universal solution, and no technology can substitute for strong political and corporate commitments to reducing emissions.
However, considering the urgency, scale, and complexity of the global energy transition, we cannot afford to disregard any tools in our arsenal. Used effectively, AI will expedite the energy transition while broadening access to energy services, fostering innovation, and ensuring a secure, resilient, and affordable clean energy system. It is time for industry stakeholders and policymakers to establish the groundwork for this AI-powered energy future and to form a trustworthy and collaborative ecosystem around AI for the energy transition.
In the energy sector, our research indicates that digital applications can contribute up to 8% of greenhouse gas (GHG) reductions by 2050. This could be accomplished by improving efficiency in carbon-intensive processes and enhancing energy efficiency in buildings, as well as by utilizing artificial intelligence powered by cloud computing and highly networked facilities with 5G to deploy and manage renewable energy.
An excellent example of this is IntenCity – the Schneider Electric building is equipped with IoT-enabled solutions, creating an end-to-end digital architecture that captures more than 60,000 data points every 10 minutes. It is smart-grid ready and energy-autonomous, featuring 4,000 m2 of photovoltaic panels and two vertical wind turbines.
IntenCity has its own building information modeling system, which is an accurate representation of the construction and energy model capable of replicating the energy behavior of the actual building.
In the materials sector, digital applications can lead to up to 7% of GHG reductions by 2050. This would be achieved by enhancing mining and upstream production and leveraging foundational technologies such as big data analytics and cloud/edge computing. Furthermore, use cases leveraging blockchain could enhance process efficiency and promote circularity.
In mobility, digital applications could reduce up to 5% of GHG emissions by 2050, according to our research. This would involve utilizing sensing technologies like IoT, imaging, and geo-location to gather real-time data for informing system decision-making, ultimately improving route optimization and reducing emissions in both rail and road transport.
For instance, Mobility-as-a-Service (MaaS) platforms are increasingly serving as advanced mobility planning tools for consumers, offering a wide range of low-carbon options such as eBikes, scooters, or transit.
Uber has incorporated non-rideshare options into its customer app and digital platform, utilizing analytics to suggest transportation solutions for consumers. Other studies have shown an estimated emission reduction of over 50% if MaaS could replace individual private car use.
There are high-priority, impactful use cases that, if scaled, can deliver the most benefits in the energy, materials, and mobility sectors.
The opportunity is evident: companies can expedite their net-zero goals by adopting digital use cases with high potential for decarbonizing industries. While many World Economic Forum partner companies are beginning to implement such pioneering examples, they can learn from each other and collaborate to swiftly transform their businesses, systems, workforces, and partnerships on a wide scale.
First, businesses must ensure that their data is shared, autonomous, connected, and allows for transparency to support various outcomes – from identifying and tracing source materials to optimizing routes and enhancing efficiency. They must invest in new data architectures and integrate recognized frameworks into their internal reporting structures. This ensures that data is available, standardized, and shareable across value chains and with partners outside their traditional operating environment.
Second, businesses must prioritize digital inclusion and skills development. They must ensure that their current and future workforce has access to new technologies and the necessary skills to scale digital technologies and transform business processes in high-emission industries.
Third, businesses must foster collaboration among their digital, sustainability, and operations teams, not only within their enterprises but also across value chains and industries. Partnerships between private companies, startups, technology providers, investors, and public agencies will be crucial for scaling investments , reducing the risks associated with technologies, and accelerating the sharing of knowledge.
Power consumption of training GPT-3
It is crucial to ensure that the digital transformations that expedite the clean energy transition are inclusive and sustainable so that the benefits are accessible to all. Furthermore, we must mitigate the emissions caused by the electrification and digitalization of industries through technological advancement and the development of supportive policies.
In an ever-changing time characterized by constant change, the convergence of AI and sustainable development represents a glimmer of hope, ready to redefine our joint response to pressing global issues. As environmental worries continue to grow, the need to speed up our journey towards sustainable development becomes more pressing. At this critical juncture, we see AI not just as an impressive piece of technology, but as a potent catalyst for positive change.
The potential of AI lies in its capacity to utilize data, streamline processes, and ignite innovation, positioning it to become an essential foundation in our shared pursuit of global advancement. Standing at the crossroads of innovation and sustainability, the need for action is mounting to transition towards a future characterized by resilience, sustainability, and mutual prosperity.
Calculating the energy consumption of a single Balenciaga pope in terms of watts and joules is quite challenging. However, we do have some insight into the actual energy cost of AI.
It’s widely known that machine learning requires a substantial amount of energy. The AI models powering email summaries, chatbots, and various videos are responsible for significant energy consumption, measured in megawatts per hour. Yet, the precise cost remains uncertain, with estimates considered incomplete and contingent due to the variability of machine learning models and their configurations.
Additionally, the companies best positioned to provide accurate energy cost information, such as Meta, Microsoft, and OpenAI, have not shared relevant data. While Microsoft is investing in methodologies to quantify the energy use and carbon impact of AI, OpenAI and Meta have not responded to requests for comment.
One key factor to consider is the disparity between the energy consumption during model training and its deployment to users. Training a large language model like GPT-3, for instance, is estimated to consume just under 1,300 megawatt hours (MWh) of electricity, equivalent to the annual power consumption of 130 US homes.
To put this into perspective, streaming an hour of Netflix requires around 0.8 kWh (0.0008 MWh) of electricity. This means you would need to watch 1,625,000 hours of Netflix to match the power consumption of training GPT-3.
However, it’s challenging to determine how these figures apply to current state-of-the-art systems, as energy consumption could be influenced by the increasing size of AI models and potential efforts by companies to improve energy efficiency.
According to Sasha Luccioni, a researcher at Hugging Face, the challenge of estimating up-to-date energy costs is exacerbated by the increased secrecy surrounding AI as it has become more profitable. Companies have become more guarded about details of their training regimes and the specifics of their latest models, such as ChatGPT and GPT-4.
Luccioni suggests that this secrecy is partly driven by competition and an attempt to deflect criticism, especially regarding the energy use of frivolous AI applications. She also highlights the lack of transparency in energy usage statistics for AI, especially in comparison to the wastefulness of cryptocurrency.
It’s important to note that training a model is only part of the energy consumption picture. After creation, the model is deployed for inference, and last December, Luccioni and her colleagues published the first estimates of inference energy usage for various AI models.
Luccioni and her team conducted tests on 88 different models across various applications, such as answering questions, object identification, and image generation. For each task, they performed the test 1,000 times and estimated the energy usage. Most tasks required a small amount of energy, for instance, 0.002 kWh for classifying written samples and 0.047 kWh for generating text. To put it in perspective, this is equivalent to the energy consumed while watching nine seconds or 3.5 minutes of Netflix, respectively, for each task performed 1,000 times.
The energy consumption was notably higher for image-generation models, averaging at 2.907 kWh per 1,000 inferences. As noted in the paper, the average energy usage of a smartphone for charging is 0.012 kWh. This means that generating a single image using AI can consume almost as much energy as charging a smartphone.
It’s important to note that these figures may not apply universally across all use cases. The researchers tested ten different systems, ranging from small models producing 64 x 64 pixel pictures to larger ones generating 4K images, resulting in a wide range of values. Additionally, the researchers used standardized hardware to facilitate a better comparison of different AI models. However, this may not accurately reflect real-world deployment, where software and hardware are often optimized for energy efficiency.
Luccioni emphasized that these figures do not represent every use case, but they provide a starting point for understanding the energy costs. The study offers valuable relative data, showing that AI models require more power to generate output compared to classifying input. Moreover, it demonstrates that tasks involving imagery are more energy-intensive than those involving text. Luccioni expressed that while the contingent nature of the data can be frustrating, it tells a story in itself, indicating the significant energy cost associated with the generative AI revolution.
Determining the energy cost of generating a single Balenciaga pope is challenging due to the multitude of variables involved. However, there are alternative approaches to better understand the planetary cost. One such approach is taken by Alex de Vries, a PhD candidate at VU Amsterdam, who has utilized Nvidia GPUs to estimate the global energy usage of the AI sector. According to de Vries, by 2027, the AI sector could consume between 85 to 134 terawatt hours annually, approximately equivalent to the annual energy demand of the Netherlands.
AI electricity consumption could potentially represent half a percent of global electricity consumption by 2027
De Vries emphasizes the significance of these numbers, stating that AI electricity consumption could potentially represent half a percent of global electricity consumption by 2027. A recent report by the International Energy Agency also offers similar estimates, suggesting a significant increase in electricity usage by data centers in the near future due to the demands of AI and cryptocurrency. The report indicates that current data center energy usage stands at around 460 terawatt hours in 2022 and could increase to between 620 and 1,050 TWh in 2026, equivalent to the energy demands of Sweden or Germany, respectively.
De Vries notes the importance of contextualizing these figures, highlighting that data center energy usage remained fairly stable between 2010 and 2018, accounting for around 1 to 2 percent of global consumption. Despite an increase in demand over this period, hardware efficiency improved, effectively offsetting the increase.
His concern is that AI may face different challenges due to the trend of companies simply increasing the size of models and using more data for any task. De Vries warns that this dynamic could be detrimental to efficiency, as it creates an incentive for continually adding more computational resources. He also expresses uncertainty about whether efficiency gains will balance out the increasing demand and usage, lamenting the lack of available data but emphasizing the need to address the situation.
Some AI-involved companies argue that the technology itself could help tackle these issues. Priest from Microsoft claims that AI could be a powerful tool for advancing sustainability solutions and stresses that Microsoft is working towards specific sustainability goals. However, Luccioni points out that the goals of one company may not fully address the industry-wide demand, suggesting the need for alternative approaches.
Luccioni suggests introducing energy star ratings for AI models, allowing consumers to compare energy efficiency similar to how they do for appliances. De Vries advocates for a more fundamental approach, questioning the necessity of using AI for certain tasks, considering its limitations. He emphasizes the importance of not wasting time and resources by using AI inappropriately.
Reducing the power consumption of hardware will decrease the energy consumption of artificial intelligence. However, transparency regarding its carbon footprint is still necessary.
In the late 1990s, some computer scientists realized they were heading towards a crisis. Manufacturers of computer chips had been increasing computer power by adding more and smaller digital switches called transistors onto processing cores and running them at higher speeds. However, increasing speeds would have made the energy consumption of central processing units unsustainable.
To address this, manufacturers shifted their approach by adding multiple processing cores to chips, which provided more energy-efficient performance gains. The release of the first mainstream multicore computer processor by IBM in 2001 marked a significant milestone, leading other chipmakers to follow suit. Multicore chips facilitated progress in computing, enabling today’s laptops and smartphones.
Now, some computer scientists believe the field is confronting another challenge due to the growing adoption of energy-intensive artificial intelligence. Generative AI can perform various tasks, but the underlying machine-learning models consume significant amounts of energy.
The energy required to train and operate these models could pose challenges for the environment and the advancement of machine learning. Wang emphasizes the importance of reducing power consumption to avoid halting development. Schwartz also expresses concerns about AI becoming accessible only to a few due to the resources and power required to train generative AI models.
Amidst this potential crisis, many hardware designers see an opportunity to redesign computer chips to enhance energy efficiency. This would not only enable AI to function more efficiently in data centers but also allow for more AI tasks to be performed on personal devices, where battery life is often critical. However, researchers will need to demonstrate significant benefits to persuade the industry to embrace such substantial architectural changes.
According to the International Energy Agency (IEA), data centers consumed 1.65 billion gigajoules of electricity in 2022, which is approximately 2% of global demand. The widespread use of AI is expected to further increase electricity consumption. By 2026, the agency predicts that energy consumption by data centers will have risen by 35% to 128%, equivalent to adding the annual energy consumption of Sweden at the lower estimate or Germany at the higher estimate.
The shift to AI-powered web searches is one potential factor driving this increase. While it’s difficult to determine the exact energy consumption of current AI algorithms, the IEA states that a typical request to the chatbot ChatGPT uses 10 kilojoules, which is about ten times more than a conventional Google search.
Despite the significant energy costs, companies view these expenses as a worthwhile investment. Google’s 2024 environmental report revealed a 48% increase in carbon emissions over 5 years. In May, Microsoft president Brad Smith stated that the company’s emissions had increased by 30% since 2020. Companies developing AI models prioritize achieving the best results, often at the expense of energy efficiency. Naresh Shanbhag, a computer engineer at the University of Illinois Urbana–Champaign, notes, “Usually people don’t care about energy efficiency when you’re training the world’s largest model.”
The high energy consumption associated with training and operating AI models is largely due to their reliance on large databases and the cost of moving data between computing and memory, and within and between chips. According to Subhasish Mitra, a computer scientist at Stanford University in California, up to 90% of the energy used in training large AI models is spent on accessing memory.
For instance, a machine-learning model that identifies fruits in photographs is trained by exposing the model to numerous example images, requiring the repeated movement of large amounts of data in and out of memory. Similarly, natural language processing models are not created by programming English grammar rules; instead, some models are trained by exposing them to a significant portion of English-language material on the Internet. This extensive training process necessitates moving substantial amounts of data in and out of thousands of graphics processing units (GPUs).
The current design of computing systems, with separate processing and memory units, is not well-suited for this extensive data movement. Mitra states, “The biggest problem is the memory wall.”
Addressing the challenge
GPUs are widely used for developing AI models. William Dally, chief scientist at Nvidia in Santa Clara, California, mentions that the company has improved the performance-per-watt of its GPUs by 4,000-fold over the past decade. Although Nvidia continues to develop specialized circuits called accelerators for AI calculations, Dally believes that significant architectural changes are not imminent. “I think GPUs are here to stay.”
Introducing new materials, processes, and designs into a semiconductor industry projected to reach a value of US$1 trillion by 2030 is a complex and time-consuming process. To encourage companies like Nvidia to take risks, researchers will need to demonstrate substantial benefits. However, some researchers believe that significant changes are necessary.
They argue that GPUs will not be able to provide sufficient efficiency improvements to address the growing energy consumption of AI and are working on high-performance technologies that could be ready in the coming years. Shanbhag notes, “There are many start-ups and semiconductor companies exploring alternate options.” These new architectures are likely to first appear in smartphones, laptops, and wearable devices, where the benefits of new technology, such as the ability to fine-tune AI models using localized, personal data, are most apparent, and where the energy needs of AI are most limiting.
While computing may seem abstract, there are physical forces at play. Whenever electrons move through chips, some energy is dissipated as heat. Shanbhag is one of the early developers of an architecture that aims to minimize this energy wastage.
Referred to as computing in memory, these methods involve techniques such as integrating a memory island within a computing core, which reduces energy consumption by shortening data travel distances. Researchers are also experimenting with various computing approaches, such as executing certain operations within the memory itself.
To function in the energy-limited environment of a portable device, some computer scientists are exploring what might seem like a significant step backward: analog computing. Unlike digital devices that have been synonymous with computing since the mid-twentieth century and operate in a clear world of on or off, represented as 1s and 0s, analog devices work with the in-between, enabling them to store more data in a given area due to their access to a range of states. This results in more computing power from a given chip area.
Analog states in a device could be different forms of a crystal in a phase-change memory cell or a continuum of charge levels in a resistive wire. As the difference between analog states can be smaller than that between the widely separated 1 and 0, it requires less energy to switch between them. According to Intel’s Wang, “Analog has higher energy efficiency.”
The drawback is that analog computing is noisy and lacks the signal clarity that makes digital computation robust. Wang mentions that AI models known as neural networks are inherently tolerant to a certain level of error, and he’s researching how to balance this trade-off. Some teams are focusing on digital in-memory computing, which circumvents this issue but may not offer the energy advantages of analog approaches.
Naveen Verma, an electrical engineer at Princeton University and the founder and CEO of start-up firm EnCharge AI, anticipates that early applications for in-memory computing will be in laptops. EnCharge AI’s chips utilize static random-access memory (SRAM), which uses crossed metal wires as capacitors to store data in the form of different amounts of charge. According to Verma, SRAM can be manufactured on silicon chips using existing processes.
These analog chips can run machine-learning algorithms at 150 tera operations per second (TOPS) per watt, compared to 24 TOPS per watt by an equivalent Nvidia chip performing a similar task. Verma expects the energy efficiency metric of his technology to triple to about 650 TOPS per watt by upgrading to a semiconductor process technology that can trace finer chip features.
Larger companies are also investigating in-memory computing. In 2023, IBM detailed an early analog AI chip capable of performing matrix multiplication at 12.4 TOPS per watt. Dally states that Nvidia researchers have also explored in-memory computing, although he warns that gains in energy efficiency may not be as significant as they seem. While these systems may consume less power for matrix multiplications, the energy cost of converting data from digital to analog and other overheads diminishes these gains at the system level. “I haven’t seen any idea that would make it substantially better,” Dally remarks.
IBM’s Burns concurs that the energy cost of digital-to-analog conversion is a major challenge. He suggests that the key is determining whether the data should remain in analog form when transferred between parts of the chip or if it’s better to transfer them in 1s and 0s. “What happens if we try to stay in analog as much as possible?” he asks.
Wang remarks that several years ago he wouldn’t have anticipated such rapid progress in this field. However, he now anticipates that start-up firms will bring in-memory computing chips to the market in the next few years.
The AI-energy challenge has also spurred advancements in photonics. Data transmission is more efficient when encoded in light compared to along electrical wires, which is why optical fibers are used to deliver high-speed Internet to neighborhoods and connect banks of servers in data centers. Although bringing these connections onto chips has been difficult, optical devices have historically been bulky and sensitive to small temperature variations.
In 2022, Stanford University’s electrical engineer Jelena Vuckovic developed a silicon waveguide for optical data transmission between chips. Losses during electronic data transmission are approximately one picojoule per bit of data, while for optics, it’s less than 100 femtojoules per bit. Vuckovic’s device can transmit data at a given speed for about 10% of the energy cost of doing so electronically. The optical waveguide can also carry data on 400 channels by leveraging 100 different wavelengths of light and utilizing optical interference to create four modes of transmission.
Vuckovic suggests that in the near future, optical waveguides could offer more energy-efficient connections between GPUs, potentially reaching speeds of 10 terabytes per second. Some scientists are considering using optics not only for data transmission but also for computation. In April, engineer Lu Fang and her team at Tsinghua University in Beijing introduced a photonic AI chip that they claim can produce music in the style of Johann Sebastian Bach and images in the style of Edvard Munch while using less energy compared to a GPU.
Zhihao Xu, a member of Fang’s lab, referred to this system as the first optical AI system capable of handling large-scale general-purpose intelligence computing. Named Taichi, this system can deliver 160 TOPS per watt, representing a significant improvement in energy efficiency compared to a GPU, according to Xu.
Fang’s team is currently working on making the system smaller, as it currently occupies about one square metre. However, Vuckovic anticipates that progress in all-optical AI may be hindered by the challenge of converting large amounts of electronic data into optical versions, which would involve its own energy cost and could be unfeasible.
Mitra from Stanford envisions a computing system where all the memory and computing are integrated on the same chip. While today’s chips are mostly planar, Mitra predicts that chips consisting of 3D stacked computing and memory layers will be achievable. These would be based on emerging materials that can be layered, such as carbon-nanotube circuits. The closer physical proximity between memory and computing elements offers approximately 10–15% improvements in energy use, but Mitra believes that this can be significantly increased.
The major obstacle to 3D stacking is the need to change the chip fabrication process, which Mitra acknowledges is quite challenging. Currently, chips are predominantly made of silicon at extremely high temperatures. However, 3D chips, as envisioned by Mitra, should be manufactured under milder conditions to prevent damaging the underlying layers during the building process.
Mitra’s team has demonstrated the feasibility of this concept by layering a chip based on carbon nanotubes and resistive RAM on top of a silicon chip. The initial device, presented in 2023, matches the performance and power requirements of an equivalent silicon-based chip.
Running small, ‘cheap’ models multiple times
Significant reduction in energy consumption will require close collaboration between hardware and software engineers. One energy-saving approach involves rapidly deactivating unused memory regions to prevent power leakage, and reactivating them when needed. Mitra has observed substantial benefits when his team collaborates closely with programmers. For example, by considering that writing to a memory cell in their device consumes more energy than reading from it, they designed a training algorithm that resulted in a 340-times improvement in system-level energy delay product, an efficiency metric that accounts for both energy consumption and execution speed. “In the old model, the algorithms people don’t need to know anything about the hardware,” says Mitra. That’s no longer the case.
Raghavendra Selvan, a machine-learning researcher at the University of Copenhagen, believes that there will be a convergence where chips become more efficient and powerful, and models become more efficient and less resource-intensive.
Regarding model training, programmers could adopt a more selective approach. Instead of continuously training models on large datasets, programmers might achieve better results by training on smaller, tailored databases, resulting in energy savings and potentially better models.
Schwartz is investigating the possibility of conserving energy by running small, ‘cheap’ models multiple times instead of running an expensive one once. His group at Hebrew University has observed some benefits from this approach when using a large language model to generate code. “If it generates ten outputs, and one of them passes, you’re better off running the smaller model than the larger one,” he says.
Selvan, the creator of CarbonTracker, a tool for predicting the carbon footprint of deep-learning models, urges computer scientists to consider the overall costs of AI. Like Schwartz, he believes that there are simple solutions unrelated to advanced chip technologies. For instance, companies could schedule AI training runs when renewable energy sources are being used.
The support of companies utilizing this technology will be essential in addressing the issue. If AI chips become more energy efficient, they may end up being used more frequently. To prevent this, some researchers advocate for increased transparency from the companies responsible for machine-learning models. Schwartz notes that there is a lack of information regarding the size and training data of these models.
Sasha Luccioni, an AI researcher and climate lead at the US firm Hugging Face in Montreal, Canada, emphasizes the need for model developers to disclose details about how AI models are trained, their energy consumption, and the algorithms used when a user interacts with a search engine or natural language tool. She stresses the importance of enforcing transparency.
Schwartz points out that between 2018 and 2022, the computational expenses for training machine-learning models increased tenfold every year. Mitra states that following the current trajectory will lead to negative outcomes, but also highlights the immense opportunities available.
Electricity currently constitutes between 40% and 60% of the expenses associated with data center infrastructure, and the energy requirements driven by generative AI are anticipated to increase significantly over the coming years.
The intense demand for generative AI (genAI) platforms is leading to a substantial rise in the deployment of energy-hungry GPUs and TPUs in data centers, with some operations expanding from tens of thousands to over 100,000 units per server farm.
As cloud computing and genAI gain traction, new data centers are expanding in size. It is becoming common to see new facilities designed with capacities ranging from 100 to 1,000 megawatts — which is roughly equivalent to the power needs of between 80,000 and 800,000 households, as reported by the Electric Power Research Institute (EPRI).
Energy consumption related to AI is predicted to rise approximately 45% over the next three years. For instance, the widely used chatbot, OpenAI’s ChatGPT, is estimated to consume around 227 million kilowatt-hours of electricity each year to manage 78 billion user requests.
To illustrate, the amount of energy that ChatGPT uses in a single year could supply power to 21,602 homes in the U.S., based on research by BestBrokers, an online service that analyzes trading odds derived from big data. “While this represents just 0.02% of the 131 million U.S. households, it remains a significant figure, especially considering that the U.S. ranks third globally in terms of household numbers,” BestBrokers stated in a recent report.
GenAI models generally consume far more energy than applications focused on data retrieval, streaming, and communications — the primary drivers of data center expansion over the past twenty years, according to EPRI’s findings.
At 2.9 watt-hours per ChatGPT request, AI queries are estimated to utilize ten times the energy of traditional Google searches, which consume around 0.3 watt-hours each; and the emerging computation-intensive functions like image, audio, and video generation lack any prior comparisons, according to EPRI.
Currently, there are nearly 3,000 data centers in the U.S., and this number is projected to double by 2030. Although genAI applications are estimated to consume only 10% to 20% of data center electricity at present, that figure is swiftly increasing. “Data centers are expected to account for 4.6% to 9.1% of U.S. electricity generation annually by 2030, compared to an estimated 4% today,” stated EPRI.
No crisis exists at this moment — but energy needs are on the rise
While data center energy consumption is projected to double by 2028, according to research director Sean Graham at IDC, it still represents a minor fraction of overall energy consumption — just 18%. “Therefore, it’s not entirely accurate to attribute energy usage solely to AI,” he stated. “This isn’t to suggest that AI isn’t consuming a substantial amount of energy and that data centers aren’t expanding rapidly. Data center energy usage is increasing at a rate of 20% annually. That’s noteworthy, but it still constitutes only 2.5% of global energy demand.
“It’s not as if we can lay the energy issues entirely at AI’s feet,” said Graham. “It is a problem, but AI conveniently serves as a scapegoat for the energy challenges faced globally.”
Each GPU in an AI data center can draw over 400 watts of power while training a single large language model (LLM) — which serves as the algorithmic foundation for genAI tools and platforms. As a result, merely training one LLM like ChatGPT-3 can lead to power consumption of up to 10 gigawatt-hours (GWh). This amount is roughly equal to the yearly electrical consumption of more than 1,000 U.S. homes.
“Interestingly, training the GPT-4 model, which has a staggering 1 trillion parameters, used an astonishing 62.3 million kWh of electricity over a span of 100 days,” noted BestBroker’s report. “This is 48 times greater than the energy consumed by GPT-3, which, in comparison, required about 1.3 million kWh in just 34 days.”
There are hundreds of such data centers worldwide, primarily operated by major tech companies such as Amazon, Microsoft, and Google, according to a University of Washington study. Furthermore, the energy consumption of these centers is increasing rapidly. In 2022, the total energy used by AI data centers in the U.S. reached 23 trillion terawatt-hours (TWh). (A TWh signifies one trillion watts of energy utilized for one hour.)
This figure is expected to grow at a combined annual growth rate of 44.7% and will likely reach 146.2 TWh by 2027, as per IDC Research. By that time, AI data center energy consumption is predicted to account for 18% of total data center energy use.
Given the rapid emergence of genAI, there is speculation that a crisis may arise sooner rather than later. Tech entrepreneur Elon Musk remarked earlier this year that by 2025, there may not be enough energy to sustain the swift advancements in AI.
A billing system with two levels?
In addition to the pressures from the growth of generative AI, electricity costs are increasing due to supply and demand factors, environmental regulations, geopolitical events, and extreme weather conditions driven partly by climate change, as stated in a recent IDC study. IDC believes that the elevated electricity prices observed over the past five years are likely to persist, significantly increasing the operational costs for data centers. (Building a data center can cost between $6 million and $14 million for every megawatt, and IDC indicates the typical lifespan of each center is between 15 to 20 years.)
In light of this context, electricity providers and other utilities have suggested that AI developers and operators should be obligated to pay more for electricity—similar to what cloud service providers faced earlier—due to their rapidly growing consumption of computing resources and energy relative to other users.
These suppliers further claim that they need to enhance their energy infrastructure to accommodate the heightened demand. For instance, American Electric Power (AEP) in Ohio has proposed that owners of AI data centers commit to a decade-long agreement to cover at least 90% of the energy they project they’ll require each month, even if their actual usage is lower. AEP has projected a load increase of 15 GW from data centers by 2030 and seeks upfront funding to expand its power facilities.
Data center operators, predictably, are resisting this proposal. Currently, Google, Amazon, Microsoft, and Meta are contesting AEP’s suggestion. Last month, these companies argued before Ohio’s Public Utilities Commission that such specialized rates would be “discriminatory” and “unreasonable.”
Graham refrained from commenting on whether special electricity rates for AI providers would be just, but he cited the precedent of offering lower rates for bulk industrial power consumers. “When considering the average consumer—regardless of market nuances—one might expect discounts for larger quantities,” he noted. “Therefore, data center operators likely anticipate similar volume discounts.”
Electricity constitutes the primary expense in data center operations, comprising 40% to 60% of infrastructure costs, Graham explained; altering this cost structure could have a “significant effect” on corporate profitability.
Even semiconductor manufacturers are observing the scenario with caution. Concerned about the rising power demands, Nvidia, Intel, and AMD are all developing processors designed to use less energy as a strategy to mitigate the issue. Intel, for instance, plans to soon introduce its upcoming generation of AI accelerators, shifting its emphasis from traditional computing and memory capabilities to power consumption per chip.
Nuclear energy as a potential solution.
Meanwhile, AI data center operators are exploring an unconventional energy source: nuclear power. Earlier this year, Amazon invested $650 million to acquire a data center from Tesla that operates entirely on nuclear energy sourced from one of the largest nuclear power plants in the United States.
Additionally, just last week, Microsoft revealed it is in negotiations with Constellation Energy to revive the Three Mile Island power facility in Pennsylvania—site of the most severe nuclear disaster in US history. Through this agreement, Microsoft would secure all the power generated from Three Mile Island for the following two decades to support its substantial energy requirements for AI.
In July, the US Energy Advisory Board published a report outlining strategies for supplying power to AI and data centers, offering 16 suggestions on how the US Department of Energy can assist in meeting the rising demand reliably and affordably. The report examines power requirements for AI model training, operational flexibility for data center and utility operators, and promising technologies for energy generation and storage to accommodate load increases.
Within the report, the agency mentioned that electricity providers, data center clients, and other significant consumers had consistently expressed concerns regarding their capacity to meet demand, with “almost unanimous recommendations” to expedite the addition of generation and storage, postpone retirements, and invest more in existing resources.
These recommendations include “upgrading and renewing permits for existing nuclear and hydroelectric facilities,” as well as rapidly demonstrating new clean, reliable, cost-effective, dispatchable technologies. “In many cases, [stakeholders] view the addition of new natural gas capacity—as well as solar, wind, and battery options—as key strategies available today to ensure reliability,” the report indicated.
“We will require all energy sources, including geothermal and hydrogen,” stated Graham from IDC. “The demand for power in AI is genuinely increasing. There are certain parallels that can be drawn with cloud computing, but one distinguishing feature of AI is the sheer scale of energy consumption per server.”