Artificial Intelligence also seeks its 'green' software: the battle to lower its costs is also fought in algorithms

Artificial intelligence faces a paradox: the more it is used, the more expensive it becomes to maintain. The expansion of data centers, increasingly larger models, and the arrival of agents are skyrocketing computing needs and, with them, electricity consumption. Until now, much of the solution has been sought in hardware, but software can also make a significant difference.

“Hardware consumes because it has software running,” explains Coral Calero, professor of Languages and Computer Systems at the University of Castilla-La Mancha. The researcher thus summarizes one of the principles of so-called green software: the better a program utilizes machine resources, the less unnecessary work it will have to do, and the lower its energy consumption will be.

One of the avenues researchers are studying involves reducing the precision of certain calculations. Verónica Bolón Canedo, an AI researcher at the University of A Coruña and director of the UDC-INDITEX Chair of AI in Green Algorithms, explains that some operations can go from using 64 bits to 16 or 8 bits. “It's like losing decimals,” she points out, but this simplification allows for a reduction in necessary resources.

The problem is that not all applications can afford the same margin of error. A movie recommendation can tolerate some imprecision; a system used to support a medical diagnosis cannot. Therefore, the goal is not simply to use fewer bits, but to modify the algorithm to compensate for the loss of precision and keep the result within acceptable limits.

Efficiency can also be achieved by ensuring that a model does not have to use its full capacity in every query. Mixture of Experts (MoE) architectures allow only certain parts of a neural network to be activated depending on the task. The emergence of DeepSeek once again placed this technique at the center of the debate on how to build capable models without forcing the entire network to work constantly.

Training an advanced model requires enormous amounts of computation, but the expense doesn't end when the system is ready. Every time a user makes a query, another small workload begins: inference.

And the problem multiplies with AI agents, capable of dividing a task into several steps, consulting information, using tools, and reprocessing the results. “Companies are interested in models being efficient not because they are especially concerned about the environment but because it will not be economically viable for them,” points out Calero.The difference can be considerable. A study published in Joule by Microsoft researchers calculates a median of 0.31 Wh per optimized query, but warns that requests requiring long reasoning can multiply consumption by more than ten. The work also points to a wide margin for improvement: combining advances in the models themselves, in the software that runs them, and in the hardware could reduce the energy needed for a query by between 8 and 20 times.

The pressure is not only energetic. It is also beginning to be felt in company accounts. McKinsey's State of AI 2026 report indicates that around a fifth of organizations limit the use of AI due to its operating costs, including those associated with token consumption. In parallel, another analysis by the consultancy warns that inference will become the main type of AI workload in data centers before 2030 and could exceed 40% of total demand. This completely changes priorities. Optimizing a model so that it needs fewer resources each time it responds can end up being as important as reducing the cost of training it.

The dimension of the problem helps to understand why efficiency has become a priority. According to the International Energy Agency (IEA), data centers consumed around 415 TWh of electricity in 2024, approximately 1.5% of global electricity consumption. The figure corresponds to all data centers, not just AI systems, but the expansion of accelerated computing—driven primarily by artificial intelligence—is one of the major drivers of this growth. In its central scenario, the IEA estimates that electricity consumption by data centers could exceed 945 TWh in 2030, close to 3% of global demand.

That's why optimizing software can have a cumulative effect. A small saving in a single query may seem irrelevant, but when a model processes millions of requests a day, the difference multiplies. Efficiency also allows for reducing pressure on data centers themselves and on the networks that have to power them.However, the bill for artificial intelligence doesn't end at the plug. A review published in Nature Reviews Clean Technology warns that the emissions associated with manufacturing chips and building the data centers that make this technology possible must also be accounted for. In the case of large AI data centers, this embodied footprint can represent more than half of their emissions. Software efficiency, therefore, can not only serve to reduce the electricity needed to respond to a query: if it allows more to be done with less computing capacity, it can also help contain the need for new infrastructure and the environmental costs associated with it.

The artificial intelligence race thus enters a new phase. Companies continue to seek more capable, reliable, and secure models, but they increasingly have more incentives to make these models also cheap to run. It will not always be necessary to resort to the most powerful system: many tasks can be solved with smaller models or by activating only a part of the architecture.

The objective that summarizes this new trend is simple: do the same work with less computation. And there, software can become a piece as important as chips. As Calero reminds us, measuring consumption and optimizing resources allows addressing sustainability from the very design of applications, a line of research that is also being promoted from the UCLM.There are already results that show that this saving can be achieved from the software itself. Research published in 2026 tested an automatic optimization system on Llama 3.2 and managed to reduce the energy consumption of the 1-billion-parameter model by 13.5% in certain tests, in addition to cutting its execution time. The result does not mean that any model will consume 13.5% less, but it does demonstrate that optimizing the way an AI is executed can translate into measurable savings.AI needs to continue growing, but it is becoming increasingly evident that it will not be able to do so solely by adding power. The next big leap may be in making algorithms intelligent enough to know when they don't need to use all that power.