Hello from the Buckeye state.

The walls are closing in on Narendra Modi, as US President Donald Trump pressures the Indian prime minister to stop feeding oil into an economy propped up on cheap Russian crude. My colleagues Andres Schipani, Krishn Kaushik and Anastasia Stognei outline the bind Modi finds himself in, caught between the two old cold war foes.

And trouble continues at BP, which expects to axe 6,200 jobs as it launches its second business review in six months. The cuts are part of a broader makeover for the company, which has slashed spending on clean energy and increased its oil production.

Finally, Trump’s sacking of Bureau of Labor Statistics commissioner Erika McEntarfer was just the latest blow to an agency that says funding and staff cuts have hampered its ability to collate vital reports on the world’s largest economy.

Today we take a deep dive into data centre efficiency, which experts say could mitigate the need for massive grid upgrades and additions. — Martha

Here is a much discussed problem in the US build-out of energy infrastructure to power artificial intelligence.

Booming demand for AI and cloud computing is driving the construction of power-hungry data centres. The sprawling campuses that “hyperscalers” such as Meta, Google and Amazon are building will either siphon power from the grid or generate power on-site.

In both cases, the play is to supply more energy to meet growing demand.

But that requires massive capital investment (which critics say will raise costs for households and businesses); the expansion of reliable but polluting energy sources such as gas; bets on moonshot technologies including small modular nuclear reactors; and wading through a labyrinthine permitting process.

Some researchers and companies think there’s another way: boosting data centre efficiency, thus reducing the need to build out the grid.

Data centre efficiency refers to the amount of energy that goes into powering computing equipment, versus what’s lost to cooling systems, lighting and other supporting infrastructure.

While hyperscalers have always had an incentive to care about data centre efficiency, Steven Carlini, Schneider Electric’s vice-president of innovation and data centre solutions, says it’s becoming a bigger priority as tech companies battle to win the AI race while balancing strained energy resources.

“There’s a limited amount of available power, but the more efficiently they can use that power, the more capacity they can build,” he said.

Innovations in technologies such as liquid cooling are helping to achieve these aims.

Chips generate heat as they cycle through the trillions of calculations that power large language models and image recognition. Until recently, air cooling systems were used to mitigate this, but they are becoming less economical as AI becomes more advanced. Since liquids are denser, liquid cooling systems are increasingly being adopted.

Hyperscalers such as Amazon are switching to use custom-made liquid cooling systems, in which a “cold plate” is placed directly on top of its chips and a liquid that absorbs and removes heat is run through. The system will be put into use this summer across Amazon’s data centres.

As data centres grow bigger and bigger — and harder to keep optimally cool — they are starting to use AI tools to tweak temperatures in real time.

“You don’t want to overcook the liquid going into the chips, so we’re always adjusting and optimising the temperatures,” said Carlini.

“That’s new because running at these types of densities, at this scale, is really challenging for the industry.”

Other efforts involve maximising the efficiency of the chips and power management hardware, driven by innovation from companies such as Nvidia and Groq.

An April paper from the Rand Corporation and Epoch AI found that across 500 “AI supercomputers” there was a year-on-year 1.6-times improvement in performance per chip from 2019 to 2025, driven primarily by the adoption of new models. The supercomputers also became more energy efficient, with computational performance per watt increasing by 1.34 times per year.

AmberSemi is building semiconductors that it says can cut energy waste by 15 per cent.

Electricity enters data centres at high voltages, which need to be stepped down to a level that the equipment can safely use. As power is transferred to the motherboard, the main circuit board inside a computer or server, as much as 50 per cent can be lost, according to chief executive Thar Casey.

The company uses “vertical power delivery” to supply chips with power from underneath, instead of moving it laterally, which helps conserve energy.

Casey says the market for data centre efficiency technologies will only grow.

“Each 1 per cent of energy loss represents about $460mn per year,” he said. “We’re talking about serious money, and that’s only in the US.”

Konstantin Pilz, an author of the Rand-Epoch AI report, says the biggest efficiency gains will come from chips, but these advancements are becoming increasingly hard to achieve as data centres demand more and more powerful semiconductors to support their AI models.

“Chips are the main thing that could improve and energy efficiency, which companies are aware of,” he said. “But they’re still very power hungry.”

Researchers at Duke University caused a stir earlier in the year with a paper on large load flexibility, the ability of data centres to reduce their power usage — for instance by switching to on-site power sources or deprioritising non-urgent tasks — at times of low demand or when the grid is under strain.

The researchers found that US grids could support at least 76 gigawatts of new load with an annual load curtailment rate of 0.25 per cent (meaning new loads are reduced for 0.25 per cent of their operating hours).

“Necessity is the mother of invention,” said report author Tyler Norris. “We have extreme supply chain constraints right now, so either [the demand] will disappear entirely or it can be more flexible.”

The primary critique of the study is that while it is theoretically workable, private companies have little incentive to lean into large load flexibility. Since AI models are not at full maturity, and hyperscalers are pouring massive amount of capex into their data centres, it might take an exacerbated power crunch for it to gain traction.

Still, Google is making moves in this direction. On Monday it announced agreements with two utilities, Indiana Michigan Power and Tennessee Valley Authority, to vary energy use at data centres in their purview.

Its an encouraging sign of hyperscaler interest, although some of the details are hazy.

If this all sounds good, there’s one rub: Jevons paradox.

This theory says that increased efficiency leads to more overall consumption of a resource, rather than less.

If data centres manage to reduce their overall power usage, companies may just build more and develop increasingly powerful models.

Pilz points out that we lack accurate economic modelling of how efficiency tracks with data centre demand, but says he’s “sceptical” that it provides significant downward pressure.

“If I’m using ChatGPT and I have the choice between the most advanced model or like a generation behind, I prefer to use the most advanced model,” he said. “And I think that’s why, despite these efficiency improvements, we still see this large increase in compute and power demand.” (Martha Muir)