Can AI Operate Efficiently? Researchers May Have Found a Way

Artificial intelligence (AI) continues to shape our future across almost every industry as its adoption grows to automate crucial business processes, reduce costs, and boost productivity. 

Data shows more than 80% of businesses are utilizing AI to some extent, seeing it as a core technology within their organizations. Financial commitment to AI has also risen as its adoption yields significant benefits and much higher ROI.

With that, the AI market size grew beyond $184 billion in 2024 and is projected to surpass $826 billion by the end of this decade, as per Statista.

All this growth and the explosion of interest in AI has resulted in an arms race to develop the technology. AI models like large language models (LLMs) that are being built by organizations use algorithms to understand, summarize, translate, and generate human language and are trained on vast amounts of data using power-intensive processors. 

For this, companies and data center operators are deploying lots of capital to build new high-capacity data centers. Investment in new data centers, as per the International Energy Agency (IEA), has already soared over the past two years, led by the US. In 2023, the total capital investment by Microsoft, Google, and Amazon alone was higher than that of the entire oil and gas industry of the country.

These data centers require a lot of electricity to power them. Back in 2022, the IEA stated that data centers consumed 1.65 billion gigajoules of electricity, which is about 2% of global demand. 

The global power demand from data centers is further forecasted to increase by 50% by 2027 and by 165% by 2030, according to Goldman Sachs Research from earlier this month. 

The current power usage of data centers is estimated to be around 55 gigawatts (GW), comprising cloud computing workloads (54%), traditional workloads like storage (32%), and AI (14%). Modeling future demand for these workload types gives projected power demand by 2027 to be 84 GW, where AI will grow to 27% while cloud drops to 50% and traditional workloads fall to 23%.

Global market capacity of data centers, meanwhile, is currently standing around 59 GW, 60% of which is provided by hyperscale cloud providers like Google and Alibaba and third-party data center operators with the rest coming from traditional corporate and telecom-owned data centers.

The power demand of these large hyperscale data centers is 100 MW or more with an annual electricity consumption of about the same as the electricity demand from as many as 400,000 electric cars.

As organizations continue to embrace AI, we can also expect to see the rise of AI-dedicated data centers, which is an emerging class of infrastructure designed specifically to serve the distinct requirements of AI workloads. While very few exist so far, they are primarily located in Asia Pacific and North America, most notably in Beijing, Shanghai, Northern Virginia, and the San Francisco Bay Area. 

As per Goldman Sachs Research estimates, by the end of 2030, there will be around 122 GW of data center capacity online. Due to the higher processing workload demands of AI, the density of power use in data centers is likely to grow as well to 176 Kilowatt-hours per square foot in two years.

With data centers contributing to a growing need for power, the electric grid will also need significant investment, which is estimated to be $720 billion of grid spending through 2030.

Then there’s the carbon dioxide emissions of data centers, which can also go on to more than double between 2022 and 2030. So, the electricity consumption of data centers, which is currently 1-2% of overall power and is likely to increase to 3-4% by the end of the decade, has become a cause of concern, though it’s not all a lost cause. 

The numbers, while massive, are modest when considered in a broader context of total electricity consumption growth globally, which is projected by IEA to grow by 6,750 TWh by 2030 driven by the rise of not only AI but also continued economic growth, EVs, ACs, and electricity-intensive manufacturing.

Additionally, not only the speed of AI use remains uncertain but continued efficiency improvements in both hardware and software can help address the power demands of AI technology. 

The efficiency of AI-related computer chips has been doubling about every 2.5-3 years. A modern AI chip now consumes 99% less power while performing the same calculations as a model from 2008. Additionally, AI models themselves are becoming more efficient. Coupled with new cooling technologies and innovative scientific solutions, these developments are reducing energy usage and minimizing their environmental impacts.

Click here to learn why traditional storage cannot keep up with AI.

Scientists Develop a Groundbreaking Solution 

Illustrative image. Not an actual photograph.

Against the backdrop of AI evolution and its growing energy demands, researchers in Japan have developed an innovative spintronic device that promises a substantial reduction in power consumption by allowing the electrical control of magnetic states. 

This breakthrough can revolutionize AI hardware by making chips that mirror the way neural networks function and are far more energy-efficient. Instead of using an electric current to store information, these devices use the spin of electrons for the same, which makes them faster and more efficient.

Spintronic devices are used in mass-storage devices and computing to increase the speed and performance of computers.

Researchers from the National Institute for Materials Science, Tohoku University, and the Japan Atomic Energy Agency are now utilizing it for low-power AI chips and demonstrated a proof-of-concept current-programmed spin-orbit torque (SOT) device.

SOT offers a promising mechanism for electrically encoding information in magnetic states. However, in existing schemes, it is passively determined by the material and device structures. 

“While spintronic research has made significant strides in controlling magnetic order electrically, most existing spintronic devices separate the role of the magnetic material to be controlled and the material providing the driving force.”

– Shunsuke Fukami, Tohoku University

Spintronic devices, once fabricated, have a fixed operation scheme under which they usually switch information in a binary manner; from “0” to “1“ fashion. The new research, however, has made advances in electrically programmable switching of multiple magnetic states. Manipulating the intrinsic SOT polarity allows for flexibly programmable SOT devices.

According to the results of the study published in Nature Communications¹ this month, this spintronic device enables the electrical mutual control of non-collinear antiferromagnets and ferromagnets for efficient switching of magnetic states. 

The electrical control of magnetic states is the foundation of magnetic memory, logic, and computing, and the SOT offers an efficient approach to controlling them. 

What it means is that the breakthrough SOT device can store and process information much like a brain-inspired AI chip while using significantly less energy, opening the doors for a new generation of AI hardware that is highly efficient as well as energy-saving.

The study used Mn3Sn, a non-collinear antiferromagnet, as the key magnetic material and applied an electrical current to it, which led it to produce a spin current that drove the switching of CoFeB, the neighboring ferromagnet. This was done through a process called the magnetic spin Hall effect.

The spin Hall effect (SHE) is a key element of modern spintronics that accomplishes “coupling between charge currents and collective spin dynamics in magnetically ordered systems.”

Previous research² from Tohoku University, in collaboration with other institutes, demonstrated that compared to non-magnetic materials, antiferromagnets have stronger spin Hall properties. Also, in Mn3Sn, the SHE has an anomalous sign change when its triangularly ordered moments change orientation. It also attributed Mn3Sn’s high magnetic mechanism to the spin splitting, which is dependent on momentum and is generated by the non-collinear magnetic order, expanding the scope of antiferromagnetic spintronics.

In the latest research, the ferromagnet responded to the spin-polarized current and on top of that, influenced the magnetic state of Mn3Sn. This enabled the electrical mutual switching between the two materials.

The team actually showed in their experiment that information written to the ferromagnet can be controlled electrically with the help of Mn3Sn’s magnetic state. They were able to switch CoFeB’s magnetization in different traces representing multiple states by adjusting the set current.

This current polarity changing the sign of the information written, which is an analog switching mechanism, is a key operation in neural networks and mimics the way analog values function in AI processing.

The discovery, according to Fukami, is a critical move forward in producing more energy-efficient AI chips. 

“By realizing the electrical mutual switching between a non-collinear antiferromagnet and a ferromagnet, we have opened new possibilities for current-programmable neural networks,“ he added, noting that their focus is now on “further reducing operating currents and increasing readout signals, which will be crucial for practical applications in AI chips.”

With data centers consuming so much energy, another study aimed to reduce power consumption with SOT Magnetic Random-Access Memory (MRAM).

This separate study³, which was published in Nature Communications last month, was from researchers of Johannes Gutenberg University Mainz (JGU), who developed a powerful and highly-efficient solution for data processing and storage while reducing energy consumption.

The prototype “paves the way for faster, more efficient memory solutions,“ said the lead study author, Dr. Rahul Gupta, who’s a former postdoctoral researcher at the JGU Institute of Physics.

It is based on SOT MRAM, which is known for its superior power efficiency, non-volatility, and performance compared to static RAM. The technology that uses electrical currents to switch magnetic states faces challenges in terms of reducing the high input current needed during the writing process. This means maintaining sufficient thermal stability to store the data for a long time and minimizing the energy required to perform the storage task.

Instead of SHE, this study utilized the Orbital Hall Effect (OHE), where a transverse flow of orbital angular momentum is produced in a material when an electric current is applied. Using this phenomenon, researchers were able to get greater energy efficiency without relying on expensive or rare materials.

The unique magnetic material developed uses elements like Ruthenium as a SOT channel to enhance performance. With that, the team was able to achieve a 20% reduction in the input current, a 30% increase in efficiency, and an over 50% reduction in overall energy consumption.

Companies Leading the Way

Now, let’s take a look at prominent names that are helping advance the AI space:

1. NVIDIA Corporation (NVDA +3.16%)

Nvidia designs and manufactures graphics processing units or GPUs. A GPU is an electronic circuit that performs mathematical calculations at high speed and is a popular choice for developing AI models.

Over the past decade, the company has significantly improved the performance-per-watt of its GPUs, as much as 4,000 times, according to its chief scientist William Dally and continues to develop specialized circuits in these chips that are specifically designed for AI computations.

Nvidia is the world’s second-largest company by a market cap of $3.21 trillion. Its shares are currently trading at $130.86, down 2.35% YTD. The company’s EPS (TTM) is 2.54 and the P/E (TTM) ratio is 51.67 while its dividend yield is 0.03%.

As for company financials, Nvidia will release that of 4th Quarter FY25 on Feb. 26. Now, for Q3, its GAAP earnings per diluted share were $0.78 and non-GAAP earnings per diluted share were $0.81.

For this quarter, Nvidia announced a record revenue of $35.1 billion, an increase of 17% from Q2 and 94% from a year ago. Its revenue from data centers accounted for $30.8 billion of it. For its Q4, it shared an outlook of $37.5 billion in revenue.

“The age of AI is in full steam, propelling a global shift to NVIDIA computing,“ said CEO and founder Jensen Huang at the time as he noted seeing “incredible“ demand for its Hopper as well as anticipation for Blackwell, which enables high-performance computing (HPC) for accelerated training, inference, and efficiency in gen AI.

Nvidia has been through a fantastic journey as its share price went from sub-$11 in Oct. 2022 to past $152 in early January 2025. Recently, it did lose some momentum after Chinese AI startup DeepSeek launched its latest AI model. The initial perception was that DeepSeek would undermine the demand for Nvidia’s high-performance GPUs but could actually be positive for Nvidia as the startup’s lower computing power needs may open up more AI opportunities, in turn, driving the biggest chipmaker’s prices upwards.

“AI is transforming every industry, company and country,“ Huang said during last quarter’s financials and further stated that enterprises are adopting agentic AI while breakthroughs in physical AI are being seen in industrial robotics. Countries, he added, have also awakened to the importance of building their own AI infrastructure.

#2. Broadcom Inc. (AVGO -0.23%)

Broadcom is among the top 10 largest companies, with a market cap of $1.1 trillion. It is involved in the designing, development, and supply of a broad range of semiconductor and infrastructure software solutions. The company has partnerships with major players like Google, Meta Platforms, and ByteDance. It is also working with OpenAI and Apple to build their own AI chips.

As of writing, Broadcom shares are trading at $235, up 1.95% YTD. The company’s EPS (TTM) is 1.28 and the P/E (TTM) ratio is 184.78 while its dividend yield is 1%.

Now, in Q4 of 2024, Broadcom had revenue of over $14 billion, an increase of 51% from the prior year period, while GAAP net income was $4.324 billion. For Q1 of fiscal year 2025, the company gave revenue guidance of about $14.6 billion.

In quarter four, the chipmaker’s semiconductor revenue hit a record at $30.1 billion, which CEO and President Hock Tan noted was driven by AI revenue of $12.2 bln. The 220% year-on-year growth in AI revenue was fueled by its XPUs.

Broadcom’s GAAP diluted EPS came in at $0.90 and non-GAAP diluted EPS was $1.42. Cash from operations meanwhile was $5.6 billion.The company also reported an increase of 11% in quarterly common stock dividend to $0.59 per share, which was the fourteenth consecutive increase. Free cash flow excluding restructuring also came in “strong“ at $21.9 billion while cash and cash equivalents were $9.348 billion.

“We see our opportunity over the next three years in AI as massive. Specific hyperscalers have begun their respective journeys to develop their own custom AI accelerators.”

– Tan said on the earnings call

He also shared that the company is developing AI chips with three large cloud customers, each of whom is expected to deploy 1 million AI chips in networked clusters by 2027.

Conclusion

As the popularity and usage of artificial intelligence continue to grow, so does their energy consumption. AI models, especially deep learning, require a significant amount of energy for computation, and conventional solutions face power efficiency limits.

Against this backdrop, research that aims to improve the energy efficiency of AI chips and minimize their environmental impacts can help the world realize the full potential of the technology while diminishing its negative effects.

The SOT-based memory and logic devices can make a big difference here by providing a low-power solution by using magnetic states for computation instead of charge-based logic. Retaining data without constant power can reduce energy costs for AI inference, enable more energy-efficient AI accelerators for real-time AI decision-making by mimicking the brain, and boost processing speed for AI tasks by integrating logic and memory.

Constant efforts like these make AI hardware and software more energy-efficient, faster, and capable of handling complex computations, in turn, driving broader AI adoption across industries while reducing its carbon footprint!

Click here to learn why 2025 is the year of AI-driven investments.