Powering up artificial intelligence

If you own top global tech stocks, time to get rid of your smug grin. Utilities are smashing all your precious AI plays, even though stocks like Nvidia, the chip powerhouse at the forefront of the AI revolution, continue to breach new all-time highs every other day. Nvidia stock is currently trading at over US$1,000 ($1,350) a share from around US$50 in mid-March 2020, a 40-fold jump over four years. On May 22, Nvidia reported first-quarter February-April 2024 revenues of US$26 billion, up 261% from US$7.2 billion a year ago, with earnings per share of US$6.12, up nearly six-fold from US$1.09 a year ago.

Despite’s Nvidia’s spectacular rise, three of the top five best-performing stocks year-to-date on the world’s main stocks barometer, S&P500 — Vistra Corp up 148%, Constellation Energy up 91.4% and NRG Energy up 59.3% — are utility stocks, while the other two, Super Micro Computer up 241% and Nvidia up 110%, are AI-related stocks. The utilities sector index is up 12.4% this year, and has risen 29.2% since its last October lows. In contrast, S&P500 has exactly mirrored the utilities index year-to-date but is up slightly less, or 28.5%, over the past 12 months. 

Over the past three months, the utilities index is up nearly 19% while S&P’s tech sector index is up 12.6%. (Over the past 10 years, the US utilities index has had a net total annualised return of over 8.2% compared to 12.5% annualised return for S&P500.) The utilities index, by the way, has 3.3% dividend yield while S&P500 has just 1.4% annual yield. Utilities are even crushing the Nasdaq100 which includes the top 100 US-listed tech stocks. QQQ ETF is up 14.5% this year and 13.2% over the past three months, but has a mere 0.65% annual dividend yield.

How boring utilities became AI plays
How did boring utilities get just so sexy? Power generation and electricity retailers are in vogue again because ironically, they are suddenly being seen as the ultimate AI plays. The AI revolution needs data centres full of server racks made by the likes of Super Micro Computer stuffed with the latest parallel processing H100 graphic chips from Nvidia that are needed by Microsoft, its affiliate OpenAI, as well as competitors like Google, Meta, Amazon, Oracle and others including Anthropic and Cohere, for training and inferencing of data to help them build chatbots like OpenAI’s ChatGPT4o, Anthropic’s Claude, Google’s Gemini or Microsoft’s Copilot. Over the next few weeks, iPhone maker Apple will announce a new partnership to supercharge its digital assistant Siri, while Amazon is expected to announce new plans to reinvigorate its assistant Alexa. The chatbots, Copilots and digital assistants will help power a new era of convenience, productivity and connectivity.

See also: ChatGPT’s US$8 tril birthday gift to Big Tech

Data centres are the backbone of cloud computing, providing the physical infrastructure and resources needed to store, process and distribute data across the Internet. They generate a ton of computing power or compute and need a lot of old-fashioned power or electricity. Computing accounts for 40% of the total electricity demand in a typical data centre, the International Energy Agency (IEA) estimates. Cooling the server racks to achieve stable processing efficiency consumes another 40%. The remaining 20% electricity demand in data centres is from peripheral IT equipment.

Just how much more electricity does the world need to switch from the pre-AI world where we were yesterday or a few months ago to the AI-centric world in a few years? Let’s start with some basics: Utilities have long been a boring business for a reason. While the developing world — China, India, Indonesia and Vietnam — are industrialising and electrifying remote areas which until recently had little or no access to power, in much of the developed world, demand for electricity has been flat to negative. Technology has brought in more energy efficiency and there is also growing awareness about energy conservation which has led to lower or flat overall consumption of electricity. These days, we all use LED light, which drives lower power use. Last year, America generated about as much electricity as it did 15 years ago, even though the US economy was about twice as big as it was back in 2009. 

Until now, the only growth element was electrification of transportation. So, as more electric vehicles, or electric trains, are rolled out, there would be a need for more electric power and slightly less fossil fuel like coal and oil. More recently, reshoring or onshoring of some higher-end manufacturing, away from China and Latin America, back to the US and setting up large chip plants in the US have become new demand drivers for electricity. Add to that all the electricity that is needed to power the new AI revolution. Let us stick with simple things that you and I do every day. We all google something. On average, a simple ChatGPT query needs nearly 10 times as much electricity to process as a Google search. 

See also: DBS, EnterpriseSG and IMDA launch gen AI programme to help up to 50,000 SMEs adopt tech

The IEA estimates a single ChatGPT query requires 2.9 watt-hours of electricity, compared with 0.3 watt-hours for a Google search. So, if instead of using Google search or Microsoft Bing, or for that matter DuckDuckGo, and start using ChatGPT or Claude or some other chatbot, you will need 10 times more electricity. If you Google a lot or like me, DuckDuckGo a lot, and you want to convert to AI search, you will need to start watching your electricity bill. Goldman Sachs estimates the overall increase in data-centre power consumption from AI to be in the order of 200 terawatt-hours per year between 2023 and 2030.

Selling utilities to buy US treasuries
Little wonder, then, that large US independent power producers or IPPs, which have flexibility to raise prices as opposed to regulated utilities counterparts which need permission to raise prices, are turning themselves into growth plays by hitching a ride with AI players Microsoft, Google, Meta Platform and Amazon, instead of appealing to investors who value income and safe long-term returns. The bounce in utilities index since late October comes in the wake of poor performance in 2022, as well as the first nine months of 2023 when the usually staid sector had flat net returns over a 21-month period. Utilities are normally geared up, so while they fared well during the near-zero interest rates for nearly a decade, a 5.25% increase in interest rates increased their borrowing costs and hit bottomlines. 

Moreover, their main investor base, older income-hungry retirees who normally love utilities for their steady 4% or so annual yield, sold them to buy short-term two-year or one-year US treasuries that were suddenly offering mouth-watering returns. Even a small deposit on iPhone maker Apple’s high-yield savings account earlier this year was paying over 5.2% per annum.

Utilities investors were yanking money out from poorer generating firms and piling it in high-yield deposits. Digital banks like Apple’s, which take just minutes to activate through a smartphone app, were among the big beneficiaries. US two-year yields have since plunged to 4.85%. And if you open an Apple savings account today, you might get a 4.5% rate. That is higher than what many utilities pay out in dividends, but you have an option to participate in high-flying utility stocks. The narrative outside America is that US stocks are being propelled by meme-like AI plays led by Nvidia. But look under the hood, you would find that the rally is incredibly broad-based. The rally in utilities is a case in point.

Data centres around the world consume just 1% to 2% of overall global power demand. That figure will likely rise to 3% to 4% of overall global electricity demand by the end of the decade, according to a recent Goldman Sachs research report. Data-centre power demand will grow 160% by 2030 as the pace of efficiency gains in electricity use slows and the AI boom accelerates, Goldman Sachs estimates. The biggest impact of power demand will be for US data centres. Data-centre workloads nearly tripled between 2015 and 2019 as the cloud computing phenomenon took hold. Through that five-year period, data centres’ demand for power was flat, or around 200 terawatt-hours per year. As more data centres came on stream, they became more efficient in how they used the power. By 2028, Goldman Sachs expects AI to represent about 19% of all data centres’ power demand. By 2030, data centres will use 8% of US power, compared with just 3% in 2022.

Citigroup forecasts American data centres could account for 10.9% of all US electricity demand by 2030, up from about 4.5% today. That means more power plants, more transmission lines and other infrastructure — and higher electricity costs in a higher-for-longer interest rate environment, as companies building all that capacity and infrastructure will demand better returns. Florida-based renewables giant NextEra Energy, which has wind and solar power plants, estimates that electricity demand from data centres alone is likely to grow 15% a year through the end of the decade. US utilities will need to invest around US$50 billion in new power generation capacity just to support data centres alone, the Goldman Sachs report notes. The US investment bank expects incremental data-centre power consumption in the US will drive around 3.3 billion cu ft (93.4 million cu m) per day of new natural gas demand by 2030, which will require new pipeline capacity to be built. In other words, the ongoing AI boom will unleash a whole new infrastructure boom in America. 

A bigger beneficiary of the energy infrastructure boom is likely to be Europe where at least US$1 trillion might be required to rebuild and expand infrastructure and power grid for AI. Europe’s power demand is forecast to grow by up to 50% over the next decade. Europe has the oldest power grid in the world. To keep new data centres electrified, the continent will require massive investments. Goldman Sachs expects nearly EUR800 billion ($1.17 trillion) in spending on transmission and distribution over the coming decade, as well as nearly EUR850 billion in investment in solar, onshore wind and offshore wind energy. 

Of the world’s more than 8,000 data centres, about 33% are in the US, 16% in Europe and 10% in China, according to the IEA. Outside China, within Asia Pacific, Japan and Australia with 8% and 7% of all global data centres are where the major players have their biggest footprint. Southeast Asia is emerging as a region which players like Nvidia, Microsoft, Amazon Web Service and Google have all identified as the next big frontier. For the region to become a bigger player would require a major power infrastructure investment cycle. If there is plenty of surplus power available, data-centre operators will come. That will help pave the way for a more robust AI ecosystem in the region. 

Assif Shameen is a technology and business writer based in North America