Artificial intelligence (AI) has taken over the entire world, and no sector is untouched by the magic of this technological advancement. Today, 77% of devices are incorporating AI tech, despite only a third of consumers thinking they are using AI platforms.
A growing number of organizations are also increasingly adopting AI for business growth, with 9 out of 10 companies supporting the tech for a competitive advantage. This usage is projected to eliminate 85 million jobs but, at the same time, create 97 million new ones.
These numbers reflect the estimates that the AI market will contribute $15.7 trillion to the global economy by 2030. The AI market size is also expected to grow by at least 120% year-over-year.
However, as AI adoption grows, so does the demand for energy to power data centers, which are extremely energy-intensive. Running and cooling thousands of servers around the clock requires massive amounts of energy. Currently, this energy is drawn from the electrical grid, which causes issues during peak demand and strains the system. According to Chevron (CVX +1.51%) CEO Mike Wirth:
“The growth and demand for power can burden a grid that is already stressed. It could add cost to consumers. Natural gas will help power the rapid growth of artificial intelligence with its insatiable demand for reliable electricity.”
In response to this demand, Chevron recently announced that it would work with GE (GE +2.46%) Verona and Engine No. 1 to generate electricity for AI by building natural gas power plants directly connected to data centers.
Currently, the data centers account for around 1% of global electricity consumption, as per a report by the International Energy Agency (IEA).
But this is just the beginning as companies spend billions of dollars to build even more data centers.
Microsoft (MSFT +0.46%) is planning to invest $80 billion in AI-enabled data centers, with half of it allocated to the US. According to Blackstone (BX +4.82%) estimates, over $1 trillion will be invested in US data centers through 2030, with Google (GOOGL -0.33%), Meta (META -0.76%), and Amazon (AMZN -0.41%) among the tech giants building data centers across the nation.
So, the energy consumption of the sector is only going to accelerate from here. Goldman Sachs (GS +2.06%) estimates data center power demand to grow by 160% in the next five years, which will naturally drive unprecedented growth in electricity usage.
As per the Electric Power Research Institute (EPRI) estimates, the energy consumption by these facilities in the US may consume between 4.6% and 9.1% of total electricity generation by the end of this decade.
Meanwhile, in Europe, specifically Germany, data centers consumed around 16 billion kWh of energy in 2020 for computing, storage, and transmission, a figure expected to rise to 22 billion kWh by 2025.
As a result, researchers are constantly exploring ways to bring down the power requirements of AI.
Bringing Down AI Energy Consumption
AI technology is playing a key role in enhancing energy efficiency across various sectors by optimizing energy consumption, predicting demand, and enabling smart grid management, ultimately leading to reduced energy waste and lower carbon emissions.
For instance, Vodafone UK used AI to lower the daily power usage of 5G radio units (RUs) by using advanced AI and machine learning-powered software solutions of Ericsson. They have already successfully lowered it by 33% at selected sites in London.
Vodafone UK implemented three main energy-saving features as part of the trial.
This includes a Radio Power Efficiency Heatmap, which uses ML to identify and rank underperforming sites for targeted efficiency improvements by creating a visual representation of all network cells. Then there’s the 4G Cell Sleep Mode Orchestration, which has developed a behavioral model of network cells to fine-tune sleep parameters to create a balance between energy savings and performance. 5G Deep Sleep meanwhile uses AI-driven predictive algorithms, reducing power usage by 70% during off-peak hours.
But what about AI’s energy consumption? Well, researchers and companies have been targeting a lot of different areas to reduce the energy consumption of this fast growing technology.
This includes optimizing AI models, utilizing energy-efficient hardware, and adopting sustainable practices like renewable energy and efficient cooling in data centers.
When it comes to optimizing AI models, different approaches can be taken such as pruning, quantization, and distillation. In model pruning, unnecessary parameters and connections are removed from a model without impacting the accuracy. Quantization deals with reducing the precision of model parameters to decrease computational requirements and energy consumption. In distillation, smaller models are trained to mimic the behavior of a larger, more complex model, which provides a more energy-efficient model with similar performance.
As MIT Technology Review noted in this year’s publication, smaller models are one of the technological breakthroughs that will define the next era. These models are trained on more focused datasets and can offer domain-specific precision, cost efficiency, and enhanced data security compared to large language models (LLMs).
GPT-4o mini from OpenAI, Phi from Microsoft, Gemini Nano from Google DeepMind, and Haiku from Anthropic’s Claude 3 are some examples of smaller versions of larger AI models from tech giants.
When it comes to hardware, the focus has been on using GPU and other specialized AI processors that are designed specifically for high-performance workloads.
Then, there are select energy-efficient servers and processors that adjust power usage based on workload demands, virtualization for better resource utilization by running multiple processes on fewer physical machines, and advanced cooling technologies to reduce the need for power-intensive air conditioning systems.
As we shared just last month, researchers from Tohoku University, the Japan Atomic Energy Agency, and the National Institute for Materials Science achieved a highly efficient and energy-saving AI hardware. The researchers used spintronic devices for low-power AI chips and showed a current-programmed spin-orbit torque (SOT) device that opens the doors for a new generation of AI hardware.
Yet another way to bring down AI energy consumption is through sustainable practices. Data centers can utilize renewable energy sources, implement energy-efficient cooling systems, and design efficient layouts to reduce energy consumption.
Companies are actively making commitments towards this with Google making progress with geothermal energy. Microsoft is another one, which is set to reopen the Three Mile Island nuclear plant to become carbon neutral by 2030.
Interestingly, AI is being used to bring energy efficiency to AI too. The technology is used to optimize energy consumption in buildings and industries, predicting when equipment needs maintenance, and reducing peak demand by leveraging real-time grid conditions.
Amidst this all, researchers from the Technical University of Munich (TUM) have developed a novel method that trains neural networks up to 100 times faster, significantly reducing the energy consumption associated with AI applications.
For this, the researchers computed the parameters directly based on probabilities rather than going with an iterative approach but the results have been so far comparable in quality to existing iterative methods.
A New Way to Reduce AI Energy Consumption Considerably
Besides the increased usage of AI, the launch of more complex AI applications in the coming years will result in a substantial surge in the demands of the data center capacity.
These applications will use up huge amounts of energy for the training of neural networks. So, the new study focuses on reducing the massive power required to train neural networks for AI.
Neural networks are used in AI to recognize images and process language among other tasks. The functioning of these networks is inspired by the way the human brain works.
The neutral network is made up of interconnected nodes called artificial neurons. Here, the input signals are weighted with specific parameters before being summed up and if the defined limit gets surpassed, the signal gets passed on to the next node.
Now, to train the network, the initial selection of parameter values is typically randomized and then gradually adjusted to improve the network predictions progressively.
This approach, of course, requires many iterations, which makes the training extremely demanding, in turn, leading to the consumption of a lot of electricity.
So, a team of researchers led by Felix Dietrich, a professor of Physics-enhanced Machine Learning, developed a new method. Dietricha is a core member of the Munich Data Science Institute (MDSI) and an associate member of the Munich Center for Machine Learning (MCML).
What this new technique does is, it makes use of probabilities rather than determining the parameters between the nodes repeatedly. This new probabilistic method makes a targeted use of values at critical locations in the training data. These are locations where large and rapid changes in values are taking place.
With this approach, the study aims to obtain energy-conserving dynamic systems from the data. Such dynamic systems change over time as per the rules. These systems can be found in financial markets and climate models.
“Our method makes it possible to determine the required parameters with minimal computing power. This can make the training of neural networks much faster and, as a result, more energy efficient. In addition, we have seen that the accuracy of the new method is comparable to that of iteratively trained networks.”
– Dietrich
This advancement has the potential to substantially lower the environmental impact of AI by decreasing the energy required for training neural networks. We could see industries deploying large-scale AI models adopting this method within the next 1 to 3 years, leading to more sustainable and cost-effective AI solutions.
Innovative Company
NVIDIA Corporation (NVDA -1.9%)
A leader in AI computing, NVIDIA develops energy-efficient GPUs and AI accelerators, continually innovating to enhance performance while reducing power consumption.
Last year at the AI Summit DC, the company reported that the energy usage of its GPUs has seen a massive 2,000X reduction in training over the last decade. During this period, it also experienced a 100,000X energy reduction in generating tokens. Improvement in efficiency has been 2,000X while a 4,000X improvement has been recorded in computation performance over the last ten years.
According to Bob Pette, Nvidia’s VP of enterprise platforms at the time:
“If cars had improved their efficiency as much as we have improved that inference performance, you could drive over 300 years on a single tank of gas. At the core of accelerated computing is sustainable computing.”
Nvidia’s Blackwell is its new flagship GPU architecture, which succeeds Hopper and is designed to enhance AI performance “in a significant way.” This platform is packed with the world’s most powerful chip, second-generation transformer engine, and fifth-generation NVLink while being extremely energy efficient for inference and training.
NVIDIA Corporation is actually a full-stack computing infrastructure company with its segments including Compute & Networking and Graphics.
Compute & Networking covers data center, networking, autonomous and electric vehicle solutions, DGX Cloud computing services, and Jetson, while the Graphics segment covers GeForce GPUs, GeForce NOW, virtual GPU software, NVIDIA RTX GPUs, and Omniverse Enterprise software.
NVIDIA Corporation (NVDA -1.9%)
With a market cap of $2.96 trillion, NVIDIA shares, as of writing this, are trading at $121.50, down 9.4% YTD. The company has an EPS (TTM) of 2.94 and a P/E (TTM) of 41.40, while its dividend yield stands at 0.03%.
Despite the drawdown NVIDIA share prices have been seeing lately, which isn’t exclusive to the chip maker, the company continues to attract investors. After all, its stock value has skyrocketed more than 1,800% since 2020, making it one of the most valuable companies in the world.
The interest in NVIDIA is especially among young investors as it overtakes Tesla (TSLA -5.16%) as the top-held stock on Robinhood (HOOD +7.7%). A vast majority (75%) of funded accounts on this commission-free investing platform are held by Millenials and Gen Z.
Robinhood CEO Vlad Tenev also believes Nvidia is going to be more important than ever, thanks to AI.
“I think AI is going to make investing more important because if control over the technology is going to be centralized in the technology companies, then you have to be an investor in those companies to benefit.”
– Tenev
When it comes to company financials, Nvidia reported Q4 ended January 26, 2025, revealing revenue of $39.3 billion, an increase of 12% and 78% from the previous quarter and a year ago, respectively.
In the data center segment, the company’s fourth-quarter revenue was a record $35.6 billion. During this period, NVIDIA was chosen to serve as a key technology partner for the $500 billion Stargate Project.
Other developments included cloud service providers AWS, CoreWeave, Google Cloud Platform (GCP), Microsoft Azure, and Oracle Cloud Infrastructure (OCI) using NVIDIA® GB200 systems to meet surging AI demand, making the NVIDIA DGX™ Cloud AI computing platform and NVIDIA NIM™ microservices available through AWS Marketplace, Cisco integrating NVIDIA Spectrum-X™ into its networking portfolio, collaborating with Verizon to integrate NVIDIA AI Enterprise, and partnering with IQVIA, Mayo Clinic, Illumina, and Arc Institute to advance drug discovery, genomics, and healthcare.
The revenue from Gaming and AI PC dropped to $2.5 bln while NVIDIA announced new GeForce RTX™ 50 Series graphics cards and laptops powered by the Blackwell architecture, launched GeForce RTX 5090 and 5080 graphics cards and introduced DLSS 4.
Besides these, revenue from professional Visualization was $511 million and $570 million was from the Automotive and Robotics division.
During this quarter, the company’s GAAP earnings per diluted share was $0.89, surging 14% from the previous quarter and up 82% from a year ago. Non-GAAP earnings per diluted share meanwhile was $0.89.
For fiscal 2025, revenue was $130.5 billion, an increase of a massive 114% from a year ago. GAAP earnings per diluted share for the period was $2.94 and non-GAAP earnings per diluted share was $2.99.
“Demand for Blackwell is amazing as reasoning AI adds another scaling law — increasing compute for training makes models smarter and increasing compute for long thinking makes the answer smarter.“
– CEO and founder Jensen Huang
So, the company has ramped up the production of its Blackwell AI supercomputers, achieving billions of dollars in sales.
For the first quarter of fiscal 2026, the company expects revenue to be $43 billion, GAAP operating expenses to be $5.2 billion, and non-GAAP other income to be about $400 million.
“AI is advancing at light speed as agentic AI and physical AI set the stage for the next wave of AI to revolutionize the largest industries.”
– Huang
Latest on NVIDIA Corporation
Call Traders Pile on Nvidia Stock Amid GTC Week
schaeffersresearch.com • March 17, 2025
Nvidia faces growth challenges despite strong AI demand
youtube.com • March 17, 2025
Legendary Trader Says Tesla, Nvidia, Apple Shares Are Ready To Rally: ‘I See What I Like To See’
benzinga.com • March 17, 2025
Quantum computing, AI stocks rise as Nvidia kicks off annual conference
reuters.com • March 17, 2025
U.S. politician makes super suspicious Nvidia stock trade
finbold.com • March 17, 2025
Hyperfine, Nvidia Partner To Make Brain MRIs More Affordable
benzinga.com • March 17, 2025
Conclusion
AI is set to revolutionize businesses across sectors, but that means an unprecedented increase in energy demand. This demand presents a big challenge, which many researchers are working to overcome.
The development of a new probabilistic training method marks a significant breakthrough in energy-efficient AI by reducing the power needs by up to 100 times. With AI-related energy consumption projected to surge as data centers expand, this advancement could substantially lower operational costs and environmental impact.
Combined with Nvidia’s focus on energy-efficient GPUs and AI accelerators, these innovations can help accelerate AI adoption and boost productivity.
Click here to learn all about investing in artificial intelligence.