Privacy & AI - Leadership Lens #2

This edition of Privacy & AI - Leadership Lens provides 4 strategic takeaways from the International Energy Agency's report "Energy and AI" and summarises the report in 10 charts with supporting data.

Key findings for leaders:

1. Talent is the primary barrier to AI scale. Despite rapid adoption growth (from 15% to nearly 40% in four years in large firms), the biggest constraint is missing AI expertise, not cost or tech.

• Leaders should prioritise capability-building and strategic upskilling across functions, especially in areas handling sensitive data or regulatory exposure.

2. AI is becoming an energy-intensive strategic asset. Training models like GPT-4 can require energy equivalent to powering 28,500 families for a day. Inference tasks (like generating images or videos) also come with high, variable energy costs.

• Executives should account for AI’s carbon footprint in ESG metrics, procurement strategy and annual reports.

3. Model choices drive both performance and energy exposure. Smaller, purpose-fit models or those using algorithmic efficiencies (like mixture-of-experts) can result in substantial energy savings. Mandating the use of the most efficient model for each specific job is crucial for managing computational costs and an indication of AI maturity.

• Senior leaders should incentivise development teams to consider model efficiency, rather than just defaulting to the largest, most energy-intensive model.

4. AI governance and skills are lagging in critical sectors. Industries like energy, utilities, and mining show 40% lower AI talent concentration compared to finance and tech. This creates potential operational risks in sectors where AI is increasingly tied to high-risk systems (eg. safety components of critical infrastructure, including management of the electricity grid).

• Boards should assess, as part of the risk management process, the potential for AI-related disruptions originating from utility and infrastructure providers

10 charts to understand the connection between energy and AI:

1. AI adoption rates are increasing, but larger firms and firms in higher-income countries tend to use AI more

• AI adoption is rapidly increasing. AI adoption rates increased from slightly over 15% in 2020 to nearly 40% in 2024.

• There is a significant gap in adoption rates between small and large firms, and this gap is widening (from 12% in 2020 to 30% in 2024). Missing expertise appearing to be a key constraint.

• AI adoption rates are higher in firms based in higher-income countries. For firms based in countries with a GDP per capita above USD 60.000 at purchasing power parity, adoption rates are nearly 10% higher in small firms and nearly 20% higher in large firms compared to the OECD average

2. Missing expertise is the dominant reason that firms do not adopt AI today, followed by privacy and legal concerns

For companies employing more than 250 employees:

• The top constraint hindering wider use of AI is the lack of expertise.

• Privacy and legal concerns also rated highly as impediments

• The high cost of AI tools or their lack of utility to the firm did not rate highly as barriers to adoption.

3. Training the largest AI model today requires a large amount of energy

• Training is a time-consuming and energy-intensive process

• It is estimated that GPT-4 was trained for around 14 weeks, which would result in a training energy demand of around 42.4 GWh (this is the daily electricity consumption of around 28.500 households in advanced economies)

• Energy consumption for training varies substantially depending on the model size and complexity and the hardware configuration

4. The electricity intensity of different GenAI tasks varies greatly

The amount of electricity consumed during the inference phase (user queries) depends on numerous factors, such as:

• Input query size and output answer length: longer input queries and output answers consume more electricity.

• Model size: larger models require are more electricity intensive

• Input and output mode: video and image generation are generally much more electricity intensive, than text generation

• Implementation of algorithmic efficiencies: different strategies are being deployed to reduce the computational intensity of inference, for example by using mixture of experts (MoE) models, saving on computation and energy costs while preserving model performance

• Degree of inference-time scaling: new models, such as OpenAI’s o1 or DeepSeek’s R1, use "inference scaling" to improve performance (in particular for reasoning tasks), which can dramatically increase the energy cost of inference.

• Hardware implementation: current state-of-the-art B200 GPU is 60% more energy efficient than the previous generation’s H100

Estimated energy consumption

• Text generation using a small LM consumes around 0.3 Wh, while, using a medium-sized LM consumes around 5 Wh.

• Image generation consumes around 1.7 Wh per task, but video generation more energy intensive, taking around 115 Wh to generate a short, relatively low-quality 6-second video

• In comparison, charging a mobile phone or laptop requires around 15 Wh and 60 Wh, respectively.

5. Model design and model choice have large impacts on electricity intensity

In general, larger models and reasoning models tend to consume more energy

• Very small LM consumed 0.1Wh for a task:

• Medium-sized LM used around 4 Wh for the same task

• Large reasoning model consumed twice as much electricity as a model of a comparable size

6. Efficiency measures for inferencing, such as batching, halve the electricity consumption per task

• Inferences are often processed through batching: grouping different independent inputs together and processing them in parallel

• Batching allows for more efficient utilisation of GPU computing capabilities that would otherwise be underutilised

7. Around 70% of the growth in electricity demand from servers between 2025 and 2030 comes from accelerated servers

• Accelerated servers are specialised servers equipped with GPUs or similar accelerator chips to enhance computing performance for specific tasks and they are important for AI training and deployment

• Electricity consumption in accelerated servers is projected to grow by 30% annually, while conventional server electricity consumption growth is slower at 9% per year.

• Accelerated servers account for almost half of the net increase in global data centre electricity consumption, while conventional servers account for only around 20%.

• Other IT equipment and infrastructure (cooling and other infrastructure) account for around 10% and 20% of the net increase, respectively

8. Energy use by data centres and cryptocurrencies have risen sharply since 2020, while devices and networks have seen slower growth

• Consuming around 360 TWh of electricity in 2023, data centres accounted for one-third of overall ICT sector electricity use, estimated at over 1.000 TWh in 2023, equivalent to 4% of global electricity use.

• Data centres have contributed most to ICT sector energy growth since 2020, increasing by over 90 TWh between 2020 and 2023.

• Telecommunication networks, including fixed and mobile access and core networks, consumed around 280 TWh, while personal computers, mobile phones and other connected devices used around 440 TWh.

• Cryptocurrencies consumed around 125 TWh in 2023 (0.5% of global electricity)

• TVs, peripherals and cable TV networks consumed around 500 TWh (2% of global electricity).

9. Based on the pipeline of announced projects, 15% of global data centre capacity under development is concentrated in the top 10 largest markets by installed capacity

• The existing infrastructure, policy frameworks and talent pools that enabled the top markets to flourish have created momentum that continues to draw development and justify investment in the expansion of supporting infrastructure

• More than 15% of data centre capacity under development globally falls within the top ten largest data centre markets by installed capacity, indicating the continued attractiveness of these hubs

10. The adoption of AI-specific skills has been slower in certain segments of the energy sector compared to other industries

• While demand for AI and digital skills is increasing in the energy industry, it is not rising as fast as in other sectors. One reason for this may be that energy employers are not yet prioritising AI and digital skills in hiring due to unclear use cases and applications of AI.

• The utilities and the oil, gas and mining sectors saw lower levels of AI skills than other sectors across 43 countries

• Between 2018 and 2024, the concentration of AI talent in utilities and oil, gas and mining was on average 40% lower than in education, financial services, professional services, and technology, information and media

Read the full report here

ABOUT ME

I'm leading the development of the AI governance programme at Informa Plc.

Previously I worked as senior privacy and AI governance consultant at White Label Consultancy. I previously worked for other data protection consulting companies.

I'm specialised in the legal and privacy challenges that AI poses to the rights of data subjects and how companies can comply with data protection regulations and use AI systems responsibly. This is also the topic of my PhD thesis.

I have an LL.M. (University of Manchester), and I'm a PhD (Bocconi University, Milano).

I'm the author of “Data Protection Law in Charts. A Visual Guide to the General Data Protection Regulation“ and "Privacy and AI". You can find the books here