Bubble or not? How energy costs are working against AI firms

HSBC’s Georges Elhedery warns AI revenues may not cover soaring compute costs.

Current revenues at AI companies may not justify the vast cost of compute. So said HSBC CEO Georges Elhedery at the Global Investment Summit for financial leaders in Hong Kong, according to CNBC.

In July, Morgan Stanley analysts reported that over the next five years global data-center capacity will increase sixfold, and by the end of 2028 the value of data centers and their equipment will reach $3 trillion.

An April report from McKinsey cites an even bigger figure: by 2030, meeting demand for AI infrastructure will require $5.2 trillion in capital expenditure. Spending on data centers serving traditional IT applications is forecast at $1.5 trillion.

Consumers are not ready to pay for this, Elhedery said, and companies will be cautious because productivity gains will not materialise within a year or two.

“It looks like five-year trends, and so the growth means that we will start to see real benefits in the form of revenue and willingness to pay for this probably later than investors expect,” he said.

General Atlantic chairman and CEO William Ford agreed:

“In the long term you will create a whole range of new industries and applications, and that will deliver benefits in the form of productivity gains, but it will take 10–20 years.”

Big tech groups Alphabet, Meta, Microsoft and Amazon have lifted their capital-expenditure forecasts to $380 billion in 2025. OpenAI signed a series of infrastructure agreements worth around $1 trillion.

Ford noted that the huge spending pouring into the AI sector reflects an understanding of the technology’s long-term impact. But you have to “pay up front for an opportunity that will emerge in the future.” In the early stages, there may be “misallocation of capital, overvaluation and irrational enthusiasm.”

“You are really betting that this will be a broad-based technology, more like railroads or electricity, which had a profound impact over time and changed the economy. But in the first few years it is very hard to predict exactly how,” the General Atlantic CEO concluded.

A bubble—or not?

Investors have lately been actively debating whether markets are overvaluing artificial intelligence.

Last week investor Ray Dalio said his personal “bubble indicator” is at a relatively high level. At the same time, Fed chair Jerome Powell described the AI boom as “different” from the dotcoms.

Magnus Grimeland, founder of Singapore venture firm Antler, believes the industry is “definitely” not in a bubble. In his view, the pace at which neural networks are being adopted in business is faster than with other technological shifts such as moving from physical servers to cloud computing.

Moreover, artificial intelligence is a “priority task” for opinion leaders, whether heads of medical institutions in India or American Fortune 500 companies.

“What distinguishes this situation from a bubble and makes it entirely different from the dotcoms is that most of the growth is backed by real revenues,” Grimeland said.

Another difference between the popularity of AI and the dotcom boom is the speed of consumer adoption.

“Think how quickly our behaviour on the internet changed, right? A year ago 100% of my searches [were in] Google. Now the figure is probably 20%,” Grimeland said.

AI projects are increasingly integrating with familiar online systems. In October OpenAI unveiled its Atlas browser with a built-in chatbot and AI assistants.

OpenAI’s big spend

OpenAI’s annual revenue exceeds $13 billion, CEO Sam Altman said on the BG2 podcast. That is a big number, but it pales beside the $1 trillion the startup plans to spend on compute infrastructure over the next decade.

Host Brad Gerstner asked Altman about this. He replied:

“First, we are getting much more. Second, Brad, if you want to sell your shares, I will find you a buyer.”

He added that there are critics who “get very excited talking about compute systems or something else and would be happy to buy shares.”

Altman acknowledged that there are scenarios that could cause problems, such as a lack of access to sufficient compute. But “revenues are growing rapidly,” he added.

“We are betting that growth will continue, and that is not only about ChatGPT. We hope to become one of the important AI services, our consumer devices business will be significant, and artificial intelligence capable of automating science will create enormous value,” the entrepreneur said.

Microsoft CEO Satya Nadella noted that OpenAI has “exceeded” all the business plans it provided to his company as an investor.

Big-ticket investments

Large corporations continue to pour tens of billions of dollars into AI initiatives despite talk of a sector bubble.

In April OpenAI agreed to raise $40 billion at a $300 billion valuation. In October the company allowed current and former employees to sell $6.6 billion of shares. The deal valued the startup at $500 billion—a record among private firms.

On November 3 the cloud-computing startup Lambda announced a multibillion-dollar agreement with Microsoft to build AI infrastructure based on tens of thousands of Nvidia chips.

“We are in the middle of perhaps the biggest technology build the world has ever seen. The industry is performing very well right now, and a lot of people are using ChatGPT, Claude and other available artificial-intelligence services,” said CEO Stephen Balaban.

On October 3 Microsoft announced on Monday $15.2 billion of investments in the UAE over four years. They include deliveries of advanced Nvidia graphics chips.

Under the agreement, the US granted the company a license to export the chips.

The company began spending in the region in 2023. The new agreement envisages $7.9 billion of investment from the start of 2026 to the end of 2029, including $5.5 billion for capital expenditure and the expansion of AI infrastructure.

Microsoft also signed a $9.7 billion deal with Australia’s IREN to provide AI cloud computing capacity. The agreement will give the company access to infrastructure built on Nvidia GB300 graphics processors.

Grimeland stressed that “huge” sums are flowing into AI-linked companies at the “wrong” valuation, but the sector’s opportunities are far greater.

Energy

Energy is one of the main sources of AI’s enormous costs. Running hundreds of thousands of graphics processors requires a constant flow of electricity. That strains the grid and pushes up utility bills, to the displeasure of end users.

How much energy will be enough for AI? No one knows—not even Altman or Nadella.

“In this particular case, supply-and-demand cycles are really impossible to predict. The biggest issue we face now is not an oversupply of compute but the ability to build [data centers] fast enough near power sources,” the Microsoft CEO said on the BG2 podcast.

Otherwise the company will end up with too many chips in the warehouse with nowhere to plug them in, he added.

Rising demand for electricity has outpaced utilities’ plans to add generation. That has led data-center developers to procure power off-grid through bespoke contracts.

“If a very cheap form of energy appears at scale any time soon, then many people will be extremely disappointed with the existing contracts they have signed,” Altman noted.

In July 2024 Bernstein Research warned of electricity shortages if demand from AI data centers continues to grow at the current pace.