
The race for leadership in artificial intelligence (AI) is seen as existential for companies, industries and nations. In the first two articles in our series, The Solutionist, we examined the global data-centre arms race, demand for power and local impacts. Now we turn to solutions. Industries ranging from tech to energy to engineering are all contributing new ideas and technologies to meet the moment. Which of these have the most potential, and how soon can it be realised?
Solutions to the issue of energy needs for AI growth broadly fit into three categories. First is improving the cooling solutions that eat up around a third of power in data centres. Second is improving the energy systems that produce and manage electricity. Third is adopting emerging techniques to tame the energy appetite of AI chips, algorithms and models. The advances are not just incremental. They are game-changers that can reduce energy and water use by as much as 50% or more.
Rethinking cooling
Data centres have typically been air-cooled but, as the power of chips has expanded to accommodate the needs of machine learning and AI, liquid-cooling solutions are stepping in. Whether a centre is air- or liquid-cooled is a useful, if imperfect, way of telling whether it is capable of supporting large-scale AI compute or not.
Liquid cooling not only allows for higher-density compute, it is also more energy efficient per unit of AI compute. Yet, although current liquid-cooling methods are more energy efficient than legacy air cooling, the scale of AI demand growth means new methods of liquid cooling to support increasing densities and consume less energy will be essential.
There’s no one-size-fits-all. You’re going to need multiple cooling types, even within a single facility
Steven Carlini, Vice President of Innovation and Data Center for Schneider Electric
In a typical air-cooled facility, about 40% of all the power the data centre uses goes to fans and chillers, not compute. The equivalent figure for liquid cooling is about 20%. As AI workloads surge—Nvidia’s H100 graphics processing units (GPUs) can draw up to 700 watts each—just two racks would demand nearly 40 kW. That already nudges past the safe ceiling of the best fan-based set-ups, which can handle no more than around 35 kW (see chart 1). Next-generation systems like Nvidia’s GB200 and GB300 racks can reach up to 150 kW per rack, far beyond what air cooling can support.
Retrofitting with new equipment is one answer. Upgrading older data centres with liquid rear-door heat exchangers, for example, can significantly improve air cooling. But true liquid cooling to support high-density compute (such as direct-to-chip technology, where the coolant absorbs heat directly through a cold-plate) requires more specialised server and plumbing designs. That makes retrofitting difficult.
A liquid-cooling method currently generating excitement is submerging servers directly into dielectric fluids. This can cut cooling power usage by almost 80% compared with conventional air-cooled systems. Samsung, a chip giant, says that it has achieved this level of reduction at its Hanam data centre site in South Korea.
Another company helping to popularise immersion cooling is Shell, an energy major. Shell’s biodegradable fluids can reduce electricity use by almost half, use 99% less water, and can cut data centres’ CO₂ emissions by up to 30%, according to the company. Shell is backing immersion cooling as “one of the single most promising solutions to transform the efficiency of data-centre cooling and overall energy use”, says Dr Selda Gunsel, its chief technology officer.

Increased efficiency in use of water matters. Conventional evaporative and air-based cooling systems for data centres consume vast amounts of water annually, often in arid regions where electricity is cheap and land is available but where water is in short supply. Large data centres can consume up to 5m gallons/day, equivalent to the daily water use of a town of 50,000 people.
The next frontier, beyond submerging servers in fluids, is on-chip cooling. In 2021 Microsoft, a technology company, prototyped micro-fluidic channels etched directly into silicon substrates, adapting a 2015 discovery by researchers at the Georgia Institute of Technology. These channels whisk heat away at the micron scale, slashing thermal resistance by 44% over cold plates and enabling tighter GPU packing densities (see chart 1). Some in the US Department of Energy reckon this technology could reach the market by 2028.
Once commercialised, this solution, working at the silicon level, could do away with the need for big cooling systems entirely. According to Pankaj Sharma, executive vice-president at Schneider Electric, a global energy-technology company, this marks a shift from component to system-level design: “It’s not just about the chips anymore. We’re now talking about integrated thermal management at the substrate level. This is the next frontier for AI density.” Data-centre operators will employ a range of solutions, and will need multiple cooling types, even within a single facility.
Perhaps the most radical idea is an effort to turn chips themselves into self-cooling engines. IBM Research has explored using phase-change materials that oscillate between solid and liquid states as temperatures rise and fall. Commercialisation remains years away, but the technology signals an era when compute bricks could become their own thermostats. Data centres would act merely as huge warehouses, using minimal power.
Immersion cooling is one of the single most promising solutions to transform the efficiency of data-centre cooling and overall energy use.
Dr Selda Gunsel, chief technology officer, Shell
All about the energy system
Cooling innovations and improvements matter, but the force of demand for compute means that the overall picture is still one of soaring AI electricity consumption. How those kilowatt hours are generated and delivered is critical.
Today’s hyperscale technology companies draw on a mix of on-site renewables, dedicated gas and nuclear power plants, and conventional grid supply. But AI’s breakneck growth is still far outpacing generation capacity. “Power has become the main limit on AI’s growth and therefore AI innovation and therefore overall American AI competitiveness,” says Varun Sivaram, chief executive of Emerald AI, a US startup.
That pressure is in turn driving hyperscalers to make important infrastructure decisions quickly before key technology parameters are fixed. Data-centre developers do not always know what emerging cooling technologies to design for, how much compute density to plan for or, even, what the exact source of future energy will be.
The advantage of uncertainty is that the ongoing innovation in cooling, and in computing itself, is a spur for new, advanced energy solutions that are cleaner, controllable and dispatchable. Many data-centre owners are looking to advanced nuclear. The International Energy Agency expects that nuclear power will supply roughly 18% of the world’s electricity for data centres by 2035—just over three times the amount of energy it is estimated to provide in 2025 —and renewables will provide roughly 43% (see chart 2).
Small modular reactors (SMRs) promise to scale up faster and to minimise meltdown and other safety risks, but they are still several years away from commercial operation at scale. Meanwhile, fusion efforts like Helion and Commonwealth Fusion Systems, based in Massachusetts, aim for commercial pilots by the early 2030s. In June Google signed a deal to buy 200 megawatts of fusion power from Commonwealth Fusion Systems.
It’s not just about the chips anymore. We’re now talking about integrated thermal management at the substrate level. This is the next frontier for AI density.
Pankaj Sharma, Executive vice president of secure power, data centres & global services at Schneider Electric
Hydrogen is another promising field for powering AI. Companies including Siemens Energy are working on hybrid hydrogen-gas turbines and modular electrolysis units to decarbonise industrial clusters. These technologies can make new turbines for data centres “hydrogen ready”, even if they use gas today. Separately, hydrogen fuel cells are being trialled as drop-in replacements for back-up diesel generators at electricity grid “edge” sites (where AI servers are near local applications).
Such new energy technologies are promising but are still years away from being commercially available at scale. In the meantime, data centres are turning to “bridge power” to meet their immediate needs. Many operators are temporarily installing gas turbines. Other operators are moving to where gas is already cheap and available.

Even then, there are no easy answers to accessing power quickly. “Putting data centres in regions where energy is abundant and underutilised offers a compelling solution to bring new compute online today, without having to wait for a grid connection tomorrow,” notes Cully Cavness, president of Crusoe Energy, a startup that is building Stargate, a giant AI data centre, at a cost of up to $500bn. But siting data centres away from the cities where most compute happens has disadvantages in terms of latency and performance. And even when developers can access cheap gas, turbine supply chains are clogged with years-long backlogs.
With constrained generation a reality for data centres, many people believe that there is mileage in grid flexibility. Companies are developing digital substation and grid orchestration platforms that allow utilities to shift and prioritise electricity flow based on compute loads and demand curves in real time. Such technologies aim to make the grid greener and smarter.
In Barcelona, for example, two companies, Telefonica, a Spanish telecoms company, and Endesa, a Spanish electricity group, are testing AI-assisted grid technologies that manage voltage and frequency hour by hour. Their smart grid nodes help to redirect surplus solar energy during low-demand afternoon periods to nearby data centres, then scale back power delivery as residential demand peaks in the evening. This can produce a 10-15% reduction in power wastage and an 8-12% improvement in utilisation of renewable energy.
Putting data centres in regions where energy is abundant and underutilised offers a compelling solution to bring new compute online today, without having to wait for a grid connection tomorrow
Cully Cavness, president, Crusoe Energy
Taming AI’s appetite
Recent AI models have expanded to hundreds of billions of parameters, and each training run can guzzle megawatt hours. As needs move from training to inference, with inference already consuming a larger share of power, compute energy demand patterns are changing. An effective way to curb AI’s energy consumption is hardware specialisation.
Google’s fifth-generation tensor processing units, for example, use 1.3 to 1.9 times less power compared with similarly sized GPUs by Nvidia - although Nvidia’s larger and more power-intensive chips achieve higher compute per watt used (see chart 4). Startups like Groq are also pursuing efficiency; its language processing units (LPUs) deliver up to 4.7 times the energy efficiency of the next-best accelerator for large language model tasks. Using its LPUs, Groq Cloud has reduced energy use for inference by as much as 80-90% compared with conventional GPU farms.
Algorithmic “pruning” is also gaining popularity. Techniques like quantisation, which reduces numerical precision from 16-bit to 8-bit formats, can maintain near identical model accuracy while cutting compute operations by half or even three-quarters.
“Efficiency as a service” platforms are also emerging to help data-centre operators optimise AI compute loads. NeuReality, an Israeli startup, is one provider; its NR1-S appliance achieves up to 15 times better energy efficiency compared with traditional CPU-centric systems and a 90% reduction in inference spending by shifting specific workloads to processors and times where it is most cost efficient. Carbon-aware computing can similarly shift workloads to times and places where energy is cleaner.
Emerald AI offers a glimpse into the future of flexible compute. Its orchestration layer redistributes AI workloads across locations and time, throttling usage in response to grid stress, and ramping up when energy is cleaner, cheaper or more abundant. The effect is that AI infrastructure becomes more actively involved in the functioning of the grid and contributes to its overall flexibility and responsiveness. “Flexibility lowers cost, improves reliability and accelerates AI competitiveness,” says Mr Sivaram.
Smarter orchestration and more flexible resources on the grid support better integration of intermittent renewables and reduces reliance on back-up fossil-fuel generation. And by building responsive, grid-aware systems, data centres can better co-exist alongside the needs of local communities.
There’s an AI for that
Solutions to the problem of energy for AI data centres are coming fast and furious. Not all offer the same impact or timeline. Cooling innovations allow for higher-density compute today, but the most advanced techniques are still in the lab. Such innovations, as well as power-system advances, like SMRs and hydrogen fuel cells, promise long-term gains, but costs are high and deployment timelines stretch over several years.
Adopting data-centre solutions at scale is not just a technical challenge but also one of cost constraints, operational inertia and pressure for near-term profits. Yet the potential gains are impressive. Immersion cooling cuts cooling power demand by anywhere between 40% and 80% if data centres are willing to do retrofits. In the case of better chips for inference, new techniques can lower energy use by up to 90%–and are already in use. The best solution overall, is compute-side improvements, which deliver the greatest energy savings at the lowest cost– while also providing flexibility resources that benefit the wider energy system.
The sustainability challenge for AI data centres may not be solved, but progress is rapid. In the face of uncertainty around technological change, flexibility may be the key ingredient–for power grids, AI models and how people themselves adapt to the new AI reality.
More from the series

The AI factory transforming your neighbourhood | The Solutionist

