• Datacentre power consumption levels are causing concern as the number of sites and increasing use of services rises
• The widespread use of generative AI (GenAI) services is driving additional demand 
• Power generation is already creaking at the seams and carbon emissions continue to rise
• Solutions range from the probable to the possible… and beyond

The astonishing speed of the global spread, and the increasing use, of generative AI (GenAI) is spurring the construction of many new datacentres, big and small and the amount of electricity required to keep those datacentres in operation is putting so much pressure on power grids that some are already close to collapse. 

Meanwhile the emissions that, 24 hours a day, contribute to climate change are getting worse as, in many parts of the world, the consumption of fossil fuels actually increases rather than declines. Old coal mines are being kept open and new ones being dug while more new licences are granted permitting the further extraction of hydrocarbon oils and natural gases. And the huge problems of what to do with the waste products of nuclear power stations and the coal and oil industries just get bigger, as do costs and the potential for very serious security breaches.

Some indication of the breadth and depth of the problem can be gleaned from Polar Capital, the London, England-headquartered specialist investment management company that runs a well-regarded Smart Energy Fund. It estimates the installed datacentre capacity for AI training will multiply from the 5GW or so consumed worldwide currently to more than 40GW by 2027 – that is more than the current annual power consumption of the entire UK. 

Meanwhile, research house Gartner estimates that GenAI could consume up to 3.5% of the world’s electricity. “AI consumes a lot of electricity and water. This negative impact should be mitigated,” noted Pieter den Hamer, VP analyst at Gartner. “Executives should be cognizant of AI’s own growing environmental footprint and take active mitigation measures. For example, they could prioritise (cloud) datacentres powered by renewable energy.”

And there are warnings from the sector itself – not just about how much energy is being consumed but access to it. Marc Ganzi, CEO at DigitalBridge, which owns and invests in digital infrastructure assets, such as datacentres, noted on the company’s first-quarter earnings call that the industry is “kind of running out of power in the next 18 to 24 months… we think we [will] run out of transmission infrastructure for power dedicated to datacentres in 24 months. 

“Power generation [is] not the issue. It’s power transmission and distribution that are constrained. Transmission grids are capacity challenged. And imagine, if you think it’s hard to get a new cell tower permitted, think about building new transmission towers or substations. There’s a lot of friction in the system around this right now. In fact, growing contribution from renewables, which is an important development, introduces additional complexities to the grid, especially as it relates to datacentres,” added Ganzi. 

Big technology companies are in the frame for much of the problem and face mass opprobrium as emissions rise and climate change accelerates. Hence, the well-publicised focus on innovative, clean-energy, alternative technology projects that, whilst attractive and very necessary, are in many cases likely to take many years to become reality, if they happen at all. In the meantime, efforts continue to make the conventional generation of electricity more efficient and less polluting.

Several viable alternatives are being explored and the world’s major AI infrastructure providers, including Alphabet (Google), Amazon, Meta and Microsoft, have promised they will have ceased to release all harmful emissions from all their businesses by 2030. They have also publicly committed to science-based, net-zero targets and massive investment in renewable energy from hydro-electrics to solar and wind power, although the solar and wind options only work when meteorological conditions are favourable so cannot be a consistent solution. In fact, estimates are that they will never be able to supply more than 40% of global energy needs, at best. 

As for practical solutions at chip level, work is progressing on the manufacture of application-specific integrated circuits (ASICs), which are designed to execute workloads with much greater efficiency while consuming less power. Elsewhere, power-efficiency gains are being enhanced by the use of next-generation semiconductor materials based on gallium nitride (GaN) and silicon carbide (SiC), which reduce power conversion losses by up to 70% in comparison to today’s conventional silicon-based components.

Meanwhile, and simultaneously, the cooling of AI data processors and the reduction of energy consumption in datacentres is becoming ever more problematic because they are running very close to the limits of conventional air-cooling technology. One important alternative is liquid cooling, as liquids have more efficient heat transfer properties than air. Options here may well include direct-to-chip liquid cooling, rack rear-door cooling and next-generation immersion cooling, where a printed circuit board is fully submerged in an electrically insulated liquid solution. According to Polar Capital, if direct-to-chip liquid cooling technology was applied to AI-training GPUs (graphics processing units), power consumption could be cut by 30%: However, as of today, just 5% of datacentres are liquid cooled.

The high-power and low-latency interconnection of many thousands of GPUs in an AI datacentre and across networks also chews up huge amounts of power and, to help ameliorate the problem, a lot of work is being done on the development of optical fibre cross-connects to replace power-hungry and hot-running electrical cross-connects. Fibre-optic technology also enhances data transmission spectrum space while significantly reducing packet loss in transmission. Another benefit of fibre is the use of advanced modulation techniques. All these solutions, and more, help keep temperatures, and power usage, down.

Geothermal, flared gas, modular mini-nuclear fission and even nuclear fusion seen as candidate solutions

At more of a macro level, the big tech firms, and many lean and lithe startups, are running feasibility studies and experimenting with geothermal, flared gas and, of course, nuclear technologies. Geothermal systems, drilled down to 20km or even further into the earth, tap directly into the heat constantly being generated naturally underground itself. This heat is free at source, infinitely available and just keeps on coming regardless of whatever weather conditions pertain on the surface. 

Flared gas, which is usually a flaming byproduct of oil production, is an option being pursued by some startups. Probably the best case in point at the moment is that of Crusoe, a Denver, Colorado-based outfit that started life as a crypto-mining business but which, a couple of years ago, changed direction and moved into AI. The company does deals with oil companies to convert the excess gas their wells produce and that would have been flared-off into electricity to power the datacentres it is building. Crusoe has, so far, built more than 100 identical, modular datacentres actually on the premises of, or immediately adjacent to, oil wells in fairly remote areas in American states, such as Colorado, North Dakota and Wyoming.

Crusoe provides a happy solution for the well owners, which usually burn off waste gas in the open air, where, because the gas never completely combusts, it produces a lot of methane, and methane warms the earth’s atmosphere. Crusoe burns the natural waste gas by stoichiometric combustion, a process in which the amount of oxygen exactly matches the amount of fuel that is to be burned. In such cases all available oxygen is completely consumed with the result that the incidence of methane is reduced by 99.9%. Crusoe says it will be capable of producing about 200MW of renewable power by the end of the year. The great majority of the company’s customers are AI startups, and AI startups are looking for between 5MW and 25MW to power their enterprises.

Of course, the big tech companies are looking at all possible (and some perhaps not-so-possible) solutions and are claiming that the sheer scale of their datacentres means they can be more efficient and cleaner than any smaller, privately owned datacentre operators. Amazon claims to have been the world’s largest corporate buyer of renewable energy since 2021, while Google is exploring the use of “advanced nuclear energy” as well as geothermal technologies.

On the nuclear power side of the equation, much is expected of so-called “mini nuclear reactors”, modular nuclear reactors small enough to be built on or close by the enormous datacentre campuses that are beginning to appear. Such plants have been under development for years now, but the best guess-timate for when they might be available commercially remains vague at “between five and 10 years”.

Then, at the outer limits of scientific conjecture, let alone of any feasible reality, is nuclear fusion, a technology that, were it ever to be developed, is touted at the infinite resource of clean energy and the ultimate solution to all the earth’s power needs – forever. Well, that’s the theory. Nuclear fusion actually powers our sun and the other stars of the universe – powering datacentres is something else. It’s not that humanity can’t make nuclear fusion work, it can, in terms of thermo-nuclear bombs where the destruction wrought by nuclear fusion could, within a few milliseconds, destroy us all. 

The hard part is slowing down that fusion, maintaining it and controlling it over any meaningful period of time to provide infinite power. That problem has defeated all nuclear scientists since 1946 and continues to do so. As yet, there is no evidence that fusion power can be sustained long enough to be of any practical use, although the quest to find a viable fusion solution continues apace, not least because fusion power plants would not create any long-lasting nuclear waste as is the case without current nuclear fission plants. All a fusion reactor would produce by way of waste would be the inert gas helium, and not very much of that. 

For the fusion of atomic nuclei, immense pressure and massively high temperatures are required (hotter than at the centre of the sun’s core) to negate the tremendous power of electrostatic repulsion and force the nuclei together. When that happens, two comparatively lightweight nuclei merge to form a single heavier nucleus. During the process, energy is released because the total mass of the now single nucleus is less than the mass of the two original nuclei: All the residual mass becomes energy.

Tricky, eh? But Microsoft reckons it could have nuclear fusion up and running by 2028. Perhaps CEO Satya Nadella knows something we don’t? Apparently, Microsoft is working with a company called Helion to test prototype fusion in a magnetised vacuum chamber and then transmuting that energy onwards as a practically useful form of green power. The initial goal is to extract more power than it takes to cause the reaction in the first place.

Where nuclear fusion is concerned, sceptics abound. One such, Peter Judge of Datacenter Dynamics, a publication and website dedicated to that subject, wrote: “Microsoft and Helion have fused hype and greenwash”, adding, “Microsoft's investment in fusion power will deliver overheated vapour, not actual electricity.” It’s hard to disagree, especially when one factors in the 2028 deadline. The chance must be close to zero, but we’ll see.

- Martyn Warwick, Editor in Chief, TelecomTV