Graphic: AI generated
The rapid development of AI has triggered a veritable boom in data centers. Companies such as OpenAI and DeepSeek are driving this revolution and the demand for high-performance servers is growing exponentially.
However, the increase in computing power is also accompanied by massive energy consumption, an issue that is leading to global discussions about infrastructure, efficiency and future investments [1].
At the same time, the question arises as to whether there is currently an overinvestment in computing power. The Chinese AI company DeepSeek, for example, has presented a model that works more efficiently than previous large language models (LLMs).
Does this mean that we will soon need less computing power?
Or will the Jevons paradox occur instead, i.e. the effect that more efficient technologies actually increase overall consumption in the long term? [2, 3]
In this article, I will focus on the key developments in the data center sector, the growing demand for energy, regional characteristics, current challenges and potential investment opportunities.
As always, the article is intended to shed light on the background to current events, provide food for thought and give impetus. The stocks mentioned do not, of course, constitute investment advice.
🤖 Data centers: the foundation of the AI revolution
The growing global demand for AI-supported software and digital applications requires powerful data centers. Goldman Sachs analysts forecast that the global demand for power from data centers will increase by 50% by 2027 and by up to 165 % could increase [1].
This chart below forecasts the energy consumption of data centers (in terawatt hours) by 2030, distinguishing between AI- and non-AI-based applications in the US and the rest of the world. Total consumption is expected to rise to over 1,000 TWh by 2030 [4].
Our analysts expect data center power consumption to increase by more than 160% by 2030
Source: [4], primary: Masanet et al. (2020), Cisco, IEA, Goldman Sachs Research
This data shows how AI applications will massively increase energy consumption. The rapid increase in the area of "US AI" and "Rest of world AI" is particularly striking.
The three main reasons for this increase are
New models such as GPT-5 or DeepSeek AI require more and more computing power. Training and operating these models requires trillions of calculations [1].
- Real-time AI applications:
Companies are integrating AI into numerous applications: from search engines to personalized financial and healthcare services.
- Cloud computing & data storage:
As digitalization progresses, the global demand for data storage and cloud services is increasing [1].
Which companies dominate the market?
On the demand side for data centers, large hyperscale cloud providers and other companies are building large language models (LLMs) that are capable of processing and understanding natural language. These models need to be trained on huge amounts of information using power-intensive processors [4].
On the supply side, hyperscale cloud companies, data center operators and asset managers are deploying large amounts of capital to build new high-capacity data centers.
These include, among others:
In addition, specialized data center providers such as Equinix $EQIX (-1,8 %) and Digital Realty $DLR (-3,48 %) as they supply physical infrastructure to the hyperscalers [6].
According to Goldman Sachs Research, demand for data center infrastructure will increasingly outstrip supply in the coming years.
The utilization rate of existing data centers is expected to rise from around 85% in 2023 to more than 95% by the end of 2026. However, the situation is expected to ease from 2027 onwards as new data centers are commissioned and demand growth driven by AI slows down (see chart below) [1].
Goldman Sachs currently estimates that the global power consumption of the data center market is around 55 gigawatts (GW). This is made up of cloud computing workloads (54%), traditional workloads such as email or data storage (32%) and AI (14%) [1].
For the future, analysts predict that electricity demand will increase to 84 GW by 2027. The share of AI is expected to grow to 27%, while the cloud share will fall to 50% and traditional workloads to 23% [1].
By the end of 2030, around 122 gigawatts (GW) of data center capacity will be online.
At this point, I asked myself as a layman how the units mentioned so far are to be understood, in my first graphic I speak of 1,000 TWh of energy consumption of all data centers by 2030 and now here is talk of 122 GW of data center capacity? In order not to go completely beyond the scope of the article, I have added a section at the very end in case some of you also feel like a layman and want to put the "units" into perspective.
... and now on with the article...
One central problem remains:
Where does all the energy come from?
⚡️Energieversorgung: Can the grid keep up?
According to estimates by Goldman Sachs, more than 720 billion US dollars will have to be invested in expanding the power grid worldwide by 2030 in order to supply the new data centers with sufficient energy [1].
Europe in particular, where electricity consumption was expected to decline for many years, is experiencing a veritable "demand shock" [1].
Which energy sources supply data centers?
- Natural gas & battery storage:
Natural gas is seen as a realistic short-term solution to meet continuous demand. It serves as a bridging technology until renewable energy and storage solutions are further developed, as renewable energy is not available around the clock [4].
Wind and solar energy could cover around 80 % of demand in the long term, provided that sufficient storage solutions are integrated [4].
In practice, solar plants run on average only about 6 hours per day, while wind power plants run on average 9 hours per day. There is also a daily volatility in the capacity of these sources, depending on the radiation of the sun and the strength of the wind [4].
The graph shows the fluctuations in capacity factors for wind and solar energy in the USA in 2023. The capacity factor indicates how efficiently an energy source utilizes its maximum output throughout the year.
- Wind energy (light blue line): The highest capacity factors occur in the winter months (Jan-March) and drop significantly in the summer months (Jun-Aug).
- Solar energy (dark blue line): Efficiency rises in the spring (Mar-May) and reaches its maximum in the summer months (Jun-Aug) before falling in the winter (Nov-Dec).
The graph illustrates that wind and solar energy can complement each other seasonally: While wind is more efficient in winter, solar energy provides the highest yields in summer. This shows how important a balanced energy mix is to ensure security of supply.
In addition to finding environmentally friendly energy sources to power data centers, technology providers can reduce emissions intensity through efficiency gains.
The following chart shows the development of the workload and energy consumption of data centers between 2015 and 2023. Although the workload almost tripled, energy consumption remained almost constant until 2019 thanks to efficiency gains. The efficiency gains then slowed down from 2020 onwards.
Source: [4], primary: Masanet et al. (2020), IEA, Cisco, Goldman Sachs Research
This chart supports the discussion on the Jevons paradox (see below). Efficiency gains could be offset or even exceeded in the long term by higher workloads and AI demand. This highlights the need to make data center energy sources more sustainable.
Meanwhile, governments are also becoming more supportive of nuclear energy on the whole. Switzerland is reconsidering the use of nuclear generators for its electricity supply, while nuclear energy enjoys bipartisan support in the US and the Australian opposition party has put forward plans to introduce nuclear reactors [4].
Participants at the COP28 conference at the end of 2023, an annual summit convened by the United Nations to combat climate change, agreed to triple global nuclear capacity by 2050 [4].
Nuclear energy is considered the ideal option for basic power supply as it provides a reliable and constant supply of energy.
As a result, more and more large tech companies such as Alphabet, Amazon and Microsoft are turning to small modular nuclear power plants (SMRs).
📊 Increasing efficiency & the Jevons paradox
With new technologies such as DeepSeek, AI could work more efficiently in the future. But does greater efficiency automatically mean that less computing power is required?
The Jevons paradox: More efficiency = more consumption?
The Jevons paradox describes the fact that increases in efficiency often do not lead to lower consumption, but to higher consumption overall.
-Example:
In the 19th century, more efficient steam engines did not lead to lower coal consumption; on the contrary, as the machines became cheaper and more versatile, coal consumption actually increased.
With cars: more fuel-efficient engines did not lead to less gasoline consumption, but to people driving more cars.
-Applied to AI:
As AI models become more efficient, the cost per computation decreases. This makes AI applications attractive in even more areas, which in turn leads to a higher overall demand for computing power.
🌎 Regional distribution and global expansion of data center infrastructure
Current distribution: Where are the data centers located today?
Today, most data centers are located in the Asia-Pacific region and North America. Well-known locations are:
North America:
- Northern Virginia
- San Francisco Bay Area
Asia: Beijing
- Beijing
- Shanghai
These regions are characterized by high computing power, intensive data traffic and strong demand from corporate campuses [1].
The chart also shows the historical development of data center capacity by region (North America, APAC, etc.) from 2017 to 2024. The figures illustrate how fast the infrastructure for the AI revolution is growing and underlines why the energy requirements of data centers are increasing so rapidly.
The increase in capacity from around 20 GW in 2017 to almost 60 GW in 2024 shows an enormous growth trend. This correlates directly with the increasing demand for AI applications and cloud computing.
How is supply growing?
Goldman Sachs Research estimates that global data center capacity will increase to around 122 GW by the end of 2030, as mentioned above. The share of hyperscalers and specialized operators will increase from the current 60% to around 70% [1].
The largest expansion of data centers has been recorded here in the past ten years.
The largest expansion of new data centers is planned in North America over the next five years.
📈 Investment opportunities: Some winners of the AI and data center revolution
US shares e.g.:
- Carrier Global $CARR (-1,06 %) : Precise cooling technology and air conditioning for data centers
- Vertiv Holdings $VRT (-7,34 %) : Specialist in cooling and power solutions specifically for data centers
- Brookfield Renewable Partners $BEP.UN : Leading provider of renewable energy (hydropower, solar, wind) - supply contracts (PPAs) with data centers
- ON Semiconductor $ON (-2,61 %) : Leader in chips for energy efficiency and thermal management. Solutions reduce power consumption in data centers and support AI integration
- Texas Instrumentes $TXN (-0,26 %) : Energy-saving semiconductor products used in data center servers
- Equinix $EQIX (-1,8 %) : Specialized in data center infrastructure
- Digital Realty $DLR (-3,48 %) : Provider of physical infrastructure for data centers
- IBM $IBM (-0,47 %) : Quantum computing technologies that potentially consume less energy and development of energy-efficient AI solutions
- Arista Networks $ANET (-3,99 %) : Specialist in high-speed networking products for data centers
- Nvidia $NVDA (-3,63 %) : Leader in AI GPUs, Leader in AI training market. Best choice for large AI models and data center training
- AMD $AMD (-2,29 %) Competing with Nvidia with its own AI chips, but better positioned in the AI interference market where energy efficiency and cost-effectiveness are key. The interference market will be the next most important market, perhaps even the more important one.
- Broadcom $AVGO (-3,16 %) Profits from network solutions for data centers
- Microsoft $MSFT (-1,57 %) Google $GOOGL (-2,28 %) Amazon $AMZN (-2,39 %) : The big hyperscalers investing heavily in AI and cloud
European stocks e.g.:
- Siemens Energy $ENR (-4,27 %) Important role in modernizing power grids, integrating renewable energies and improving storage solutions for data center reliability
- Schneider Electric $SU (+0,5 %) : Leader in the development of energy management and cooling technology for data centers - specialty in the automation of both systems.
- ASML $ASML (-0,72 %) : Indispensable for modern chip production
- Infineon $IFX (-1,86 %) and STMicroelectronics $STM (-2,24 %) : Leading semiconductor companies with a focus on AI applications
- RWE $RWE (+1,08 %) and Enel $ENEL (+1,01 %) Utilities that are increasingly focusing on renewable energies for data centers
Japanese stocks e.g:
- Daikin Industries $6367 (+0,12 %) World market leader in air conditioning and cooling, offers specialized cooling systems for data centers and AI-supported plant management systems to further increase efficiency
- Tokyo Electron $8035 (+0 %) : Important supplier for semiconductor manufacturing
- Mitsubishi Heavy Industries $7011 (-3,2 %) : Works on the development of new nuclear power plants to secure the energy supply
🧠 Conclusion: AI, data centers & energy as the trend of the century?
Although some analysts warn of possible overinvestment, the figures indicate that the demand for computing power and energy for AI data centers will continue to rise sharply.
- Efficiency gains from models such as DeepSeek or new chip technologies could reduce energy consumption per computer, but the Jevons paradox means that overall demand will increase because more efficient systems will be used more often.
The biggest winners are therefore:
- Semiconductor companies: They supply the AI chips needed.
- Data center operators: They build the necessary infrastructure.
- Energy suppliers: They ensure the energy supply for the AI revolution.
In the long term, these companies could be among the biggest beneficiaries of the coming decades.
👨🏽💻 How do I position myself?
Personally, I think I am well positioned with the NASDAQ 100 $CSNDX (-1,63 %) (portfolio share of 23%), as the focus is on US technology and growth stocks. The ETF complements my All-World with a stronger weighting in innovative sectors such as AI and cloud computing.
In the near future, I will also take a closer look at Daikin Industrie $6367 (+0,12 %) share in order to increase the exposure to Japan and the share price offers an entry point at first glance.
In addition, AMD $AMD (-2,29 %) has also caught my attention, the reason being its positioning in the aforementioned interference market. Most of the capital is currently flowing into the expansion of new AI models. However, as soon as these become a "commodity" and everyone uses them, most of the capital will probably flow into the interference market (market for the application of AI models).
Furthermore, I have Siemens AG $SIE (-0,45 %) with approx. 2.3% portfolio share (still growing to approx. 4%), which I also see as well positioned for the future for the following reasons (in the context of the article):
Network stability
- Develops technologies for intelligent power grids ("smart grids"), essential for integrating renewable energies into the supply of data centers.
Data center control
- Provides automation and monitoring systems that optimize the energy consumption and efficiency of data centers
Efficient building structure
- The "Smart Infrastructure" division supports data centers with energy-efficient solutions for lighting, air conditioning and building monitoring
Not directly cooling systems, but:
- offers technologies that increase the energy efficiency of cooling systems by optimizing energy flows and data analysis
What is your opinion❓
- Which companies do you have on your radar?
- Is there a threat of overinvestment or are we just at the beginning of a century revolution?
Thanks for reading! 🤝
...Said digression follows after the sources...
__________
Sources:
[1] https://www.goldmansachs.com/insights/articles/ai-to-drive-165-increase-in-data-center-power-demand-by-2030
[2] "The Coal Question"
http://digamo.free.fr/peart96.pdf
[3] https://de.m.wikipedia.org/wiki/Jevons-Paradoxon
[4] https://www.goldmansachs.com/insights/articles/is-nuclear-energy-the-answer-to-ai-data-centers-power-consumption
[5] https://www.mckinsey.com/capabilities/quantumblack/our-insights/the-state-of-ai
[6] https://www.cbre.com/insights/reports/global-data-center-trends-2024
__________
🧭 Digression: on gigawatts and terawatt hours
In order to understand the relationship between the two figures, 122 GW (gigawatts) and 1,000 TWh (terawatt hours), it is important to clarify the units and their meaning:
Refers to the current average power capacity that data centers worldwide require to function. Power (measured in GW) describes the amount of energy consumed per second. This is therefore a snapshot of energy requirements.
- 1,000 TWh (terawatt hours):
This is an indication of energy consumption over a certain period of time, in this case one year. It describes how much energy is required in total in 12 months.
The forecast of 1,000 TWh is slightly below the value resulting from the calculation. The graph shows values slightly above 1,000 TWh; according to the calculation based on 122 GW of power capacity, energy consumption should be around 1,069 TWh.
Nevertheless, general reasons for deviations may be as follows:
- Efficiency improvements: Data centers could operate more efficiently through improved cooling, optimized hardware and software and thus consume less energy.
- Peak vs. average consumption: The figure of 122 GW could reflect peak demand, while the actual average annual demand is somewhat lower.
- Adjustments to the model: It is possible that the forecast of 1,000 TWh is conservative and does not take into account all additional loads or regional differences.
This shows how much the demand for data centers and energy will increase due to AI and digitalization by 2030
__________
