The AI revolution has a power problem

In the race for AI dominance, American tech giants have the money and the chips, but their ambitions face a new obstacle: electrical power.

“The biggest problem we have right now is not an overabundance of compute, but it’s power and…the ability to do the builds fast enough and close to power,” Microsoft CEO Satya Nadella acknowledged during a recent podcast with OpenAI chief Sam Altman.

“So if you can’t do that, you risk having a bunch of chips in stock that I can’t plug in,” Nadella added.

Echoing the 1990s dotcom frenzy to build Internet infrastructure, today’s tech giants are spending unprecedented sums to build the silicon backbone of the artificial intelligence revolution.

Google, Microsoft, AWS (Amazon) and Meta (Facebook) are tapping into their massive cash reserves to spend around $400 billion in 2025 and even more in 2026 – backed for now by eager investors.

All that money has helped ease an initial bottleneck: acquiring the millions of chips needed for the race for computing power, and tech giants are ramping up their in-house processor production as they seek to chase world leader Nvidia.

These will go into the racks that fill huge data centers, which also consume huge amounts of water for cooling.

Building huge information warehouses takes an average of two years in the United States; Commissioning new high-voltage power lines takes between five and ten years.

– Energy wall –

The “hyperscalers,” as large technology companies are called in Silicon Valley, saw the energy barrier coming.

A year ago, Dominion Energy, Virginia’s largest utility, already had a data center order backlog of 40 gigawatts, the equivalent of the output of 40 nuclear reactors.

The capacity it must deploy in Virginia, the world’s largest cloud computing hub, has since increased to 47 gigawatts, the company announced recently.

Already accused of inflating household electricity bills, data centers in the United States could represent 7 to 12 percent of national consumption by 2030, compared to 4 percent today, according to various studies.

But some experts say these projections could be exaggerated.

“Utilities and technology companies are incentivized to embrace predictions of rapid growth in electricity consumption,” Jonathan Koomey, a renowned expert at UC Berkeley, warned in September.

As during the dot-com bubble of the late 1990s, “many of the data centers that are talked about, that are proposed and in some cases even announced, will never be built.”

– Emergency coal –

If the projected growth materializes, it could create a shortage of 45 gigawatts by 2028, the equivalent of the consumption of 33 million American homes, according to Morgan Stanley.

Several US utilities have already delayed the closure of coal plants, even though coal is the most climate-polluting energy source.

And natural gas, which powers 40% of the world’s data centers, according to the International Energy Agency, is enjoying a surge in popularity because it can be deployed quickly.

In the US state of Georgia, where data centers are proliferating, a utility has requested permission to install 10 gigawatts of gas-fired generators.

Some suppliers, as well as Elon Musk’s xAI startup, have rushed to buy second-hand turbines abroad to quickly boost their capabilities. Even recycling aircraft turbines, an old niche solution, is gaining ground.

“The real existential threat right now is not the degree of climate change. It’s the fact that we could lose the AI ​​arms race if we don’t have enough power,” Interior Secretary Doug Burgum said in October.

– Nuclear, solar and space? –

Tech giants are quietly downplaying their climate commitments. Google, for example, promised net zero carbon emissions by 2030, but removed that commitment from its website in June.

Companies tend to favor long-term projects.

Amazon defends the renaissance of nuclear power thanks to small modular reactors (SMR), a still experimental technology which would be easier to build than conventional reactors.

Google plans to restart a reactor in Iowa in 2029. And the Trump administration announced at the end of October an investment of 80 billion dollars to begin the construction of ten conventional reactors by 2030.

Hyperscalers are also investing heavily in solar power and battery storage, notably in California and Texas.

The Texas grid operator plans to add about 100 gigawatts of capacity by 2030 using these technologies alone.

Finally, both Elon Musk, through his Starlink program, and Google have proposed putting chips into orbit in space, powered by solar energy. Google plans to conduct testing in 2027.

bl/arp/iv/mlm