Headlines around AI data centres are dominated by the projected energy consumption and the struggle to meet the increasing demand. Bold plans are made to secure the energy supply, feeding the appetite of hyperscale computing. The industry is currently fixated on the input side of the equation.
But this focus on the tremendous energy need is overshadowing another important aspect – the waste heat generated by data centres, therefore overlooking a crucial resource that can be harvested and utilized.
By 2030, data centres are projected to consume nearly 1,000 TWh of electricity globally, more than double their consumption in 2024. In the United States alone, they are on track to devour 10% of the entire national power grid by the end of the decade.
The energy grid is constrained and on-site generation takes years
Major digital hubs like Northern Virginia, Frankfurt, Dublin and London share the same issue. The wait times for grid interconnection have ballooned to 7-10 years. The transmission infrastructure simply cannot keep pace with the exponential demand of AI. It is a serious hurdle for AI hyperscalers to add additional computing capacity.
This has led those hyperscalers to shift to an “island mode”. Operators are no longer waiting for the utility. They are building their own on-site power plants. We are seeing a surge in orders for gas turbines and gas engines to provide prime power for hyperscale locations. Right now, the only viable way for hyperscalers to secure enough power quickly is by burning natural gas on-site.
But even building on-site generation faces severe supply chain delays. Lead times for primary generators like gas turbines or gas engines are tipping the scales at three years. Hyperscalers need fast-deployable power to ensure high availability, predictable Total Cost of Ownership and early revenue generation. Waiting for additional gas turbines to clear the supply chain to meet capacity targets is a massive financial liability.
How to utilize waste heat
There are multiple options to use the available waste heat. Currently, the industry often chooses district heating. This involves piping the hot water to nearby homes or offices. It is a proven model. However, it suffers from two fatal flaws. Firstly, most data centres are built in rural areas or industrial parks, far from residential heat networks. Secondly, even where available, district heating comes along with strong seasonal demand whereas a data centre produces heat 24/7. Residential heating is required for perhaps five months of the year. During summer, the heat demand drops to near zero, forcing the operator to vent the energy.
This is where modular waste heat recovery systems come into play. Using technology based on the Organic Rankine Cycle, operators can convert the waste heat back into baseload electricity on-site. The recovered electricity is fed back into the facility to run the cooling pumps or IT load, lowering the facility’s net demand. Crucially, this works in July just as well as in January. It solves the seasonality problem that district heating suffers from.
Modular waste heat recovery: fewer gas turbines and faster revenue
The economic impact of this integration is transformative. By capturing waste heat, data centres can generate 35% extra fuel-free on-site power per asset, effectively operating with 25% less fuel consumption. Take an example of a data centre: to reach a total capacity of 260 MWe, the site originally required 15 gas turbines. By integrating a modular waste heat recovery solution, four of those turbines become obsolete. This dramatically reduces equipment requirements, delivering a 25% faster time-to-power and unlocking revenue nine months earlier, given the typical three years delivery time for gas turbines. It is megawatt-hours that you don’t have to ask the turbine or genset manufacturers with their whopping delivery times for. With an ROI of less than four years, it is the most profitable route to hyperscale capacity.
A new perspective: efficiency and circularity
For the last ten years, the industry’s response to this energy voracity has been the Power Usage Effectiveness (PUE) metric. PUE drove operators to slash cooling costs, bringing the ratio of total power to IT power closer to 1.0. But PUE is a trap. It measures efficiency, not circularity. A data centre with a perfect PUE of 1.0 still vents 100% of its electrical input as waste heat. It is a highly efficient heater that nobody is using.
The next phase of the industry will be defined by how effectively we can monetize the heat the chips produce. Until now, waste heat recovery in data centres was a theoretical exercise that failed in practice. Air-cooled servers produced exhaust streams at 30°C-35°C, not warm enough to be useful without massive heat pumps.
Artificial Intelligence has changed this. The new generation of high-performance chips demands higher energy densities, returning also higher waste energy amounts, making the heat an additional viable fuel source for modular ORC.
Using the heat instead of venting it
Hyperscalers are already embracing this shift. A data centre that consumes a gigawatt or more of power while venting the heat is rapidly becoming economically and environmentally unviable.
Waste heat recovery is a financial hedge against soaring energy costs and a way to maximize the yield of gas fired on-site generation. From a technology perspective, it is the logical optimization of the cooling infrastructure already deployed. The era of venting usable heat into the atmosphere is over.
