TheChek News Flow

How to power data centres

China and the US are the clear frontrunners in data centre rollouts as they battle for AI leadership

The build-out of data centres is booming, spurred by the development and use of AI. China and the US are home to developers of most of the world’s foundation models and account for a significant majority of global data centre capacity.

This is only set to grow. By 2030, worldwide installed data centre capacity will reach 220 gigawatts, six times as much as in 2020, according to McKinsey forecasts. With this expansion comes a need for power — to run the chips that perform computing tasks as well as to cool the banks of servers that house them. The energy consumed by a single data centre is considerable, typically equivalent to 100,000 households, but the largest under construction are expected to be as much as 20 times bigger, according to the International Energy Agency.

Currently the largest at 150 megawatts, a data centre in Hohhot, Inner Mongolia, will be overtaken by several gigawatt-producing facilities. A project in Texas funded by Fermi America plans to use 11GW of power from a 17GW private energy grid currently under construction — although as of mid-May 2026 the one-year-old company had not signed a tenant for the campus. 

There is little evidence of scaleback, however, and collectively data centre power demand is already meaningful worldwide. In 2024, the IEA said that data centres accounted for 1.5 per cent of global electricity consumption, having grown at 15 per cent a year over the prior five years. The agency forecast that data centre consumption would more than double from 415 terawatt hours in 2024 to 945 TWh by 2030. These figures may already be out of date, as plans for data centres have in most regions outpaced forecasts. McKinsey puts global data centres’ power demand even higher, at 1,400 TWh by 2030, or 4 per cent of the world’s consumption. 

Such appetite for power is problematic, not least because the distribution of data facilities is far from even, and not always matched to energy supply — a critical bottleneck which will lend an advantage to whomever is able to resolve it.

The new technology race

China and the US are the clear frontrunners in data centre rollouts. In 2024, the two countries accounted for roughly 70 per cent of global data centre energy consumption, based on IEA figures. This dominance stems from the competition to create the leading AI models on which most other applications are built. In the US, OpenAI, Anthropic, Google and Meta are among the heavyweights investing in training or running models. DeepSeek is China’s best-known model developer but China’s leading tech companies such as Baidu and Alibaba have their own alternatives and, like the US hyperscalers, invest in the entire infrastructure stack.

Rose Luckin, professor emerita at University College London and founder of Educate Ventures Research, says: “Training a frontier model is what drives extreme data centre demand: thousands of high-end GPUs running continuously for weeks or months, at costs ranging from tens of millions to over a billion dollars per run.”

Training, however, happens “in bursts”. Inference, or the ongoing use of AI by millions of users, is a continuous compute burden that “never stops” and is only likely to grow, making it responsible for an increasing slice of the investment in data centres. 

Both the US and China would also like to increase deployment of AI in physical technology, which would eventually mean even more data centres for operation. China’s “AI+” initiative, announced in 2025, envisages the application of AI across all sectors, particularly manufacturing. Developing humanoid robots is a growing industry and a solution that China hopes might one day alleviate some of the problems associated with an ageing, shrinking population.

While the scramble to build data capacity is frequently pitched in the US as a race to beat Chinese developments, the reality may be more nuanced. Jonathan Koomey, an independent researcher formerly at Lawrence Berkeley National Laboratory who has been studying data centres for three decades, says that rapid build-out is as much a race between hyperscalers to capture demand growth and realise payback for data centre assets which are “very rapidly depreciating”. This incentive to lock in users is behind the urge to “deploy, deploy, deploy”. “The argument they are making . . . is ‘we’re gonna get there first’ . . . and the people who get there first are going to be the big winners.” The hyperscalers’ sense of urgency does not necessarily mean that society should feel the same way.

China

China’s central planning has been applied to its AI strategy with a state directive launched in 2022 summed up as “data in the east, compute in the west”. The west is relatively resource-rich, closer to power sources — many of which are renewable — and offers a cooler climate. Not only can this provide cheaper power, this approach is also less damaging to China’s climate goals. Practically, given latency constraints, many data centres still cluster closer to eastern centres of commerce, although operators including Alibaba also have facilities in Inner Mongolia and Chengdu. State-owned enterprises therefore dominate the move west. China Telecom is behind the largest data centre in Hohhot which hosts many of the country’s leading tech companies. 

US

While the Trump administration is championing AI leadership, rollout of related infrastructure is nonetheless piecemeal, dependent on private or local incentives. Construction tends to be constrained by permitting, with regulations changing between states and even at the smaller county level.

Locations such as Reno, Nevada, have attracted data centres with fast-tracked permitting, cutting months to years off the approval processes. While local authorities might be in favour, resource availability is not always suited to expansion. Conflicts over public water use and wildlife habitat issues are spurring local resistance to data centres, with billions of dollars of projects stalled or blocked. In April, lawmakers in Maine passed a bill that would have banned construction of data centres drawing more than 20MW until 2027 — a first in the US. Maine’s governor Janet Mills later vetoed the bill. But Georgia, Virginia and Oklahoma are all considering temporary bans.

Europe

Europe has lagged the US and China in data centre construction, in part because its legislated climate targets and more stringent efficiency and environmental rules can bog down approvals processes. Without a major European AI model developer or hyperscaler with its associated demand for energy, the region’s computing burden is not as heavy as for China and the US. “The issue is not that there is no demand in Europe . . . but that there is no actor willing to internalise the risk of deploying infrastructure for frontier training and inference. That’s the missing piece, and it is what the data centre footprint represents,” says Luckin at UCL.

Even without an evident homegrown AI champion, the bloc has ambitions to be a global AI player, outlined in its 2025 AI Continent Action Plan, which aims to triple EU data centre processing capacity in five to seven years. There will be challenges with this, too, given the lack of a co-ordinated energy strategy and a historic tendency to attract foreign hyperscalers. This could put a burden on resources and may not necessarily be defended as essential sovereign capacity. While countries such as Ireland once welcomed US data centres, the draw on its domestic electricity supply was so great that in 2021 it implemented a de facto moratorium on new grid connections, shunting greenfield facilities to other countries in the bloc. Recently, favoured nations with ready access to renewable power as well as cooler climates, such as Norway and Scotland, have also been facing local resistance as communities say data centres bring them few benefits. In March 2026 Edinburgh approved a temporary ban.

Rest of the world

Elsewhere, south-east Asia is becoming more concerned about data sovereignty, consequently pursuing more ambitious plans to build their own facilities. Given space constraints, Singapore also previously restricted data centre construction, instead working closely with Malaysia to expand capacity. In March, the city-state reopened for data centre applications with stricter environmental standards required for approval. All told, the IEA expects the region’s data centre energy demand to double between 2024 and 2030. 

India is another market that has grown more quickly than anticipated even by projections in the IEA’s 2024 report, which assessed installed data centre capacity of 5GW by 2030. Estimates now range from 7GW to 10GW as the country takes a more aggressive approach to upgrading its technology backbone, outlined in its IndiaAI Mission launched in 2024. The country aims to have sovereign compute and GPU clusters, requiring installation of far more capacity than previously envisaged.

The Middle East is also developing its data centre sector. Despite sub-optimal climate, from a cooling perspective, the region has attracted hyperscalers based on its access to abundant energy sources, offering a crucial competitive advantage. With the conflict in Iran, however, security risks could cause a setback to further growth ambitions.

Power play

China has the world’s largest grid, more than twice the installed capacity of the US, at nearly 4,000GW, according to official Chinese sources. Most years it adds on the equivalent capacity of a small nation. Although it is still adding fossil fuel generation, China is also the leader in renewables, with more than half of its capacity reliant on sources including wind and solar. In 2025, the power it generated from renewables exceeded the entire consumption of the EU states, according to official data.

Renewables are an intermittent energy source, which pose challenges for data centres that require stable, always-on power. In an attempt to overcome this challenge, China has invested in UHV lines to transfer power from generation to consumption regions, giving it one of the most integrated power markets in the world. It is also one of the more advanced markets in terms of battery storage capacity, which can further help to smooth out power supply. In 2025 its “new” energy storage capacity — including lithium-ion batteries but excluding more aged pumped hydro energy storage — reached 145GW, according to Energy Storage News, although this figure includes car batteries as well as the smaller, but growing, battery energy storage solution segment.

In the US, meanwhile, a sudden uptick in electricity demand in 2025 after years of stagnation is behind a surge in applications to build new power plants. To support the Reno data centres for instance, NV Energy is offering relatively low-cost power but expects to have to triple its base load. It has planned acres of solar farms and hundreds of miles of transmission lines in the desert.

Still, capacity expansion nationwide cannot keep up with demand. In 2025, America added 53GW in electricity generation, the largest addition in a single year since 2002, according to the US Energy Information Administration. China added more than 400GW of renewables capacity alone.

The Trump administration’s changes in policy on renewable energy, particularly offshore wind, with an overhaul of regulation aimed at boosting fossil fuels and permitting delays have complicated the rollout of green energy in the US. This has prompted the power industry to increase their focus on gas generation and spurred a renaissance in nuclear energy.

Companies are racing to restart mothballed nuclear plants such as Palisades in Michigan and Three Mile Island in Pennsylvania, which are expected to come online this year and next. Hyperscalers such as Amazon and Google are supporting the deployment of new smaller nuclear reactors by signing long-term power purchasing agreements. Private equity and the government are backing the sector, with Brookfield and Washington in a consortium with reactor designer Westinghouse to construct eight new large-scale plants.

But the construction and commissioning of new nuclear energy can take years to complete, limiting the sector’s ability to solve the immediate power crunch facing the AI industry.   

Against this backdrop, the US Energy Information Administration says that two-thirds of the planned 86GW of capacity additions for 2026 are solar and wind while 28 per cent will come from battery generation and just 7 per cent from gas. Other technologies, including nuclear, don’t even move the needle in the coming year. 

Improved connectivity could help to solve some of America’s power shortage issues. The US knows that without an upgraded grid its AI ambitions are hampered. In its AI Action Plan released last year, the White House said that the existing grid needed both stabilising and optimising, including strategies to enhance the efficiency and performance of the transmission system.

Power struggles

None of these strategies will be easy to realise as the push to boost power capacity has led to issues along the supply chain. Equipment and materials shortages are acute, causing problems both for the construction of new generation facilities and grid improvements. SynMax, a satellite and AI analytics group, told the FT in April that up to 40 per cent of projects due in 2026 alone were at risk of falling behind schedule due to anything from permitting hurdles to shortages of labour, power and equipment. More than 60 per cent of projects scheduled for next year have also yet to begin construction.

Expanding gas

As a result of the energy generation shortage, gas turbines have had a renaissance, but they are not a particularly efficient means of generation. Single cycle gas turbines have an efficiency of only 20 to 35 per cent, meaning that only that percentage of the gas input goes towards generating electricity while the rest is lost to heat. Koomey, the independent researcher, says: “Utilities will buy these and typically run them less than 10 per cent of the year, just for those emergency times, because they’re relatively cheap to buy and really expensive to operate.” Data centre operators, however, are running them “all the time”, a decision that prioritises speed over efficiency, driven by the urge for fast rollout.

While combined cycle gas turbines have almost double the efficiency, the order backlog is now so long from the surge in demand that delivery lead times last year were up to seven years, according to S&P Global. Costs to build plants have also risen as much as two-thirds, according to recent data from BloombergNEF. 

This has caused problems elsewhere, too. Asian utilities reliant on gas as a bridge in the energy transition are having to wait several years as the three companies that control two-thirds of the gas turbine market — Mitsubishi Heavy Industries, Siemens Energy and GE Vernova — cannot keep up with demand just three years after the industry seemed “dead”. 

Gridlock

Grid optimisation, meanwhile, is hampered by a shortage of transformers, which convert energy for long-distance distribution. Wood Mackenzie, the energy research and consultancy firm, in July 2025 anticipated a deficit in distribution transformers that year of 30 per cent, a compounding situation of increased demand coinciding with an ageing grid in need of replacement. Adding to the squeeze, the US imports 80 per cent of its power supply transformers and half of its distribution transformers, many of which come from China, where imports have been complicated by trade tensions. While the tariff regime has to a certain extent steered US companies to invest in domestic production facilities, these will take years to build. Related labour and materials are both tight in the domestic market.

Insufficient storage

Although installed capacity is increasing, battery energy storage solutions cannot be rolled out quickly enough or at sufficient scale to plug any significant gaps — coupled with fluctuating renewables sources, they are effective for power smoothing, but not for generation. Here, too, geopolitics muddies the waters. China dominates in storage, with 90 per cent of solutions reliant on the country’s lithium iron phosphate batteries, according to the IEA. While the US is attempting to increase capacity, prices in China are 30 per cent lower than the US and 35 per cent lower than the EU. The better news, however, is that overall pricing is making battery solutions more viable — a dynamic also present in the early years of solar equipment production, when China captured the market but cut-throat competition and overcapacity led to falling prices.

Going nuclear

Other new technologies heralded to fill the gap such as small modular nuclear reactors are still four to five years from viable deployment in the US, although one impediment — funding — has been resolved due to the direct involvement of hyperscalers. Nuclear power is a popular choice for the clean, stable energy needed for data centres, and SMRs can offer dedicated supply. Amazon has invested in emerging nuclear provider X-energy. In January TerraPower, founded by a group including Bill Gates, signed an agreement with Meta to develop up to eight reactor and energy storage system plants in the US. The company started construction on its first utility scale plant in April 2026. Still, of the 50 SMRs under development in the US, none has yet proved commercially viable and the only reactor with regulatory approval, NuScale Power, is itself mired in controversy, not scheduled to come online until 2030 at the earliest. 

Europe is also eyeing nuclear technology and in March 2026 introduced a strategy to accelerate the development and deployment of SMRs and advanced modular reactors — technologies that fit with its decarbonisation goals — but these will likewise not be online until the early 2030s.

Here, too, China has the advantage. The country’s first SMR went online in 2023 and its second undertook initial testing in early 2026, with an expected commercial operation date mid-year.

Other pinch points 

One area of Chinese dominance that is more tangential is its large market share in rare earths, where it has built up leading extraction and production techniques over the past couple of decades. While not directly applied to power generation or transmission, rare earths are frequently required for supporting equipment, for instance in the magnets used in wind turbines. China also dominates in many other grid-related components such as cabling.

The US should have an advantage in its access to chips from national champion Nvidia, which owns the technology central to the development of most AI models. Higher-end chips are more efficient, and therefore less power-hungry, and America has restricted exports of its most advanced processors to China, slowing down its ability to create leading-edge AI. This policy may inadvertently have stimulated creation of the most efficient model yet produced, DeepSeek, as its designers were forced to devise ways to operate with lower computing capabilities. Despite reports that further developments may be impossible without access to higher-end US tech, China’s advantage from cheap power could again see it devise innovative computing workarounds with bundles of older chips.

Data centres and the energy transition

AI data centre energy consumption is creating tension with climate goals. Regions with legislated climate targets, such as Europe, are constrained: they cannot simply compensate for renewable intermittency by expanding fossil-fuel generation. And while efforts are under way to upgrade the grid to facilitate shifting power around the bloc, it remains poorly integrated.

Despite the more permissive stance of US legislators regarding fossil fuel generation, demonstrated by the failure in early 2026 of a bill demanding tech companies secure independent energy sources to relieve the burden on local communities, hyperscalers face a similar bind. Major AI developers have set their own net zero commitments, limiting their ability to rely on carbon-intensive backup power. Gas, while widely considered a transition fuel, is not as environmentally friendly as perceived when such a large amount of capacity is involved. Analysis by Wired says that just 11 permitted power plants it examined could emit as many greenhouse gases as Morocco.

Data centres’ demand for stable, continuous power is also in conflict with renewables: wind and solar fluctuate and the output they provide cannot be guaranteed at the scale required for training-grade compute. Battery storage comes with its own drawbacks, outlined above.

Power demand

There could be other solutions to some of these tensions. The first is to get more efficient. The example of DeepSeek showed that a model need not be totally power hungry to train or function. The Chinese model made waves when it was released in early 2025, in large part due to the far-lower development costs associated with its more efficient architecture that drew less power to train — despite its reliance on older Nvidia chips which required energy-intensive bundling to achieve the necessary compute.

This efficiency has already given the model deployment traction outside the west, where power infrastructure is under-developed. Microsoft recently sounded alarm bells about the popularity of DeepSeek in regions such as Africa due to its open source model and lower costs. For the US to challenge the Chinese model’s appeal, Brad Smith, Microsoft’s president, said that African nations might need broader investment from international development banks or lending facilities, specifically to build data centres and subsidise electricity costs.

Rather than race to build ever larger models using existing structures, other developers could optimise architecture, coding language and hardware to create efficiencies in compute power usage. “The assumption built into conventional wisdom that you always need bigger models with more data may not be correct,” says independent researcher Koomey. “If your assumption is that we’re always scaling bigger AI models and always ploughing energy efficiency back into performance — of course we’re going to use more electricity.” If smaller models can work then instead of ploughing energy efficiency gains back into models to run more computing, the energy draw could be reduced.

More specialised small language models with lower demand on computing loads would also reduce power requirements.

A further solution might be to require data centre builders to power their facilities with renewables. This might seem at odds with the hyperscalers’ urgent race for build-out, but given equipment bottlenecks it could be no less speedy — and if all constructors faced the same constraints there would be no disadvantage.

“Maybe it takes them another year or two to build out with zero emissions power generation. The companies may care about that delay, but they conflate their urgency with society’s urgency . . . Those two things are not the same,” Koomey says. He adds: “The cheapest thing they could be building now is solar, wind and batteries . . . So whether or not you care about climate change, you should care about doing things in the cheapest way . . . Right now, the capacity to manufacture solar panels, wind turbines and batteries is more than up to the task of meeting that demand growth.”