How big tech plans to feed AI's voracious appetite for power

Hindustan Times

a day ago

America's tech giants are masters of the digital realm. Yet as they bet stupendous sums on artificial intelligence (ai), their ambitions are facing constraints in the physical world. Shortages of chips and data-centre equipment such as transformers and switching gear mean soaring prices and lengthy waits. Just as pressing is access to energy as utilities struggle to match the demands of Silicon Valley. On July 24th President Donald Trump published an 'ai action plan' which describes America's stagnating energy capacity as a threat to the country's 'ai dominance'. How is big tech coping with a worsening power crunch?
Demand is rocketing thanks to ever more ambitious AI plans by the hyperscalers—Alphabet, Amazon, Microsoft and Meta—all of which rely on data centres to run their services. On July 23rd Alphabet, the owner of Google, said it would increase its capital spending for 2025 by $10bn to $85bn, taking the expected combined total for the hyperscalers to $322bn this year, up from $125bn four years ago as they splash out on bigger and more power-hungry data centres (see chart 1). Mark Zuckerberg, Meta's boss, recently unveiled project Prometheus, a cluster of such centres in Louisiana covering an area almost the size of Manhattan.
Chart
New facilities consume more electricity than ever. A rack of servers stuffed with AI chips requires about ten times more power than a non-ai version a few years ago. A study by the Lawrence Berkeley National Laboratory found that in 2023 America's data centres used 176 terawatt-hours (twh) of electricity. That is forecast to increase to between 325twh and 580twh by 2028 (see chart 2), or 7-12% of America's total consumption, with hyperscalers accounting for about half.
The situation is further complicated by the shifting requirements of ai. Most of the computing power now trains AI models. As the technology is adopted more widely, more of it will be used for 'inference', when an ai system responds to a query. To speed up responses many in the industry argue that inference data centres need to be near where people are using the software. But available land and power is even harder to find near cities.
Faced with a power shortage, America's tech giants are turning to 'less than ideal places', says a former executive. Many of the preferred places such as North Virginia, with favourable tax regimes and proximity to high-capacity fibre-optic cables that ferry data around, are already overloaded with data centres. Yet even the new spots, such as Hillsboro, Oregon, and Columbus, Ohio, are becoming 'capped out', explains Pat Lynch, of CBRE, a property firm. Vacancies are near an all-time low and centres due for completion in 2028 are already fully booked.
Another strategy is to team up with smaller rivals. In June Google announced that it would rent data-centre capacity from CoreWeave, an ai cloud provider which has already signed a similar five-year $10bn leasing deal with Microsoft. Part of the capacity for such 'neoclouds' comes from repurposing facilities once used to mine cryptocurrencies.
Tech firms are also scouring the land for fresh sources of power. Amazon Web Services planned to buy and develop a nuclear-powered data centre from Talen Energy, an electricity generator. The deal was blocked by regulators for fear of raising locals' bills. On July 15th Google announced a $3bn deal for hydro-power from a dam in Pennsylvania. Hyperscalers are also playing more of a role in directly commissioning power projects. That not only includes striking deals directly with power firms but building generation capacity at data centres, to reduce reliance on grid connections.
A survey by Bloom Energy, a power provider, finds that data-centre bosses expect that 27% of facilities will have onsite power by 2030, whereas last year that share was only 1%. Google signed a $20bn deal in December with Intersect Power, a developer, to build a data centre and solar farm with battery storage. Some of the power for Meta's Prometheus project will come from natural gas extracted at the location.
The hyperscalers' desperation is helping cultivate novel sources of generation. Google has an agreement with Kairos Power, a startup developing small-modular reactors (SMRs), to provide nuclear power from 2030. Amazon has invested in X-energy, another SMR startup. Google and Meta have signed deals for geothermal energy, tapping the heat from the earth's crust. Microsoft is dabbling in hydrogen fuel cells as backup power for data centres.
Making the grid more flexible is another way to ensure reliable supplies of energy. Tyler Norris of Duke University says electricity systems are designed for extremes in demand. A hot and sunny morning in Texas, say, will send people rushing to the switch on air-conditioning units. If data centres agree not to use grid power at peak times by tapping batteries or using onsite generators, that can allow more to be added to the grid without over burdening it.
Data-centre operators that do this could get priority in the queue for power from the grid. xai, owned by Elon Musk, participated in a flexibility programme for its data centre in Memphis. SemiAnalysis, a research outfit, argues that this helped it get faster access to electricity. The tech giants are providing support in other ways, too. Google has teamed up with ctc Global, a cable-maker, to help utilities and states upgrade transmission lines.
A final strategy is to go abroad. Data centre capacity is set to soar in the Gulf countries, where big sovereign-wealth funds are bankrolling developments. Spain, with its abundant solar power, is another popular destination. Malaysia had been Asia's data-centre hotspot, thanks in part to cheap energy , though a surcharge for data centres which came into force on July 1st may put off the hyperscalers.
Making the right choice is crucial. Building huge data centres can run into trouble. 'Project Stargate', led by Openai, an ai startup, and SoftBank, a giant Japanese tech investor, has reportedly hit setbacks after disagreements about power providers and site selection. Peter Freed, an executive formerly at Meta and now a consultant, notes that building highly customised data centres for training models in the middle of nowhere may prove a bad idea. 'I worry about stranded-asset risk,' he says. And as no one knows what the demand for ai will be over the next two years even the most advanced ai model might struggle to give definitive advice.
To stay on top of the biggest stories in business and technology, sign up to the Bottom Line, our weekly subscriber-only newsletter.