How researchers want to cut AI’s energy demand
Quantum computing could enhance AI applications by accelerating selected mathematical operations and reducing the burden on conventional hardware.
HappyEva – stock.adobe.com
As AI moves deeper into vehicles, production and industrial software, its energy demand is becoming a strategic issue. Researchers are developing smaller models and new architectures, while quantum computing could one day make training more efficient.
The rising energy demand of artificial intelligence is becoming one of the major challenges of digital transformation – including in the automotive industry, where AI is moving into development, production, driver assistance and digital vehicle functions. Researchers are therefore working on ways to reduce the resource appetite of AI systems rather than simply feeding them with ever more computing power. In simple terms, the aim is to shrink the “stomach” of large language models without taking away their usefulness.
Two major research projects are addressing this challenge from different angles. One has recently concluded and delivered concrete results for more efficient AI systems. The other has just started and is looking at whether quantum computing could one day help reduce the computing effort required to train large AI models.
Why AI models need to become more efficient
The number of data centres is growing worldwide, while the energy demand of powerful AI systems is increasing at the same time. Large language models and complex neural networks in particular require substantial computing capacity for training and operation. This creates pressure not only on IT infrastructure, but also on companies that want to use AI without building oversized compute environments.
The ESCADE project, short for Energy-Efficient Large-Scale Artificial Intelligence for Sustainable Data Centers, has examined how this resource demand can be reduced. The project was funded by Germany’s Federal Ministry of Research and involved the German Research Center for Artificial Intelligence, DFKI, and Saarland University. Over three years, the partners investigated how AI models can be made more compact and efficient.
A central focus was the development of smaller and more specialised AI models. Instead of using complete models with billions or even trillions of parameters, the researchers extracted relevant knowledge from large foundation models and transferred it into much more compact student models. According to Sabine Janzen from the team led by DFKI professor and project manager Wolfgang Maaß, the approach produced strong results. In test runs, the student models delivered comparable performance while using up to 89 percent less energy.
The researchers see this not only as a contribution to sustainability. Smaller models could also make powerful AI more accessible to small and medium-sized companies that often lack the infrastructure required to operate large models.
What model compression can achieve
Alongside model compression, the consortium used methods from neural architecture search. In this approach, suitable network structures are identified and optimised automatically rather than being developed entirely by hand. According to the project team, the resulting image-processing models could be reduced in size by up to 90 percent. At the same time, energy consumption fell by around 40 percent.
The notable point is that accuracy did not suffer. Janzen said the team was even able to improve model accuracy with this approach. This is important because energy-efficient AI is often associated with compromises in performance. ESCADE suggests that this trade-off is not inevitable if models are designed, compressed and specialised carefully enough.
Automated network architecture search could therefore become an important tool for companies that want to build powerful AI systems with lower resource use. It may also help reduce the economic barrier to deploying AI in industrial environments, where computing costs, power consumption and operational reliability increasingly matter.
Where the approach was tested in industry
The researchers tested how these methods work beyond the laboratory together with SHS – Stahl-Holding-Saar. The partners developed an AI system for automated scrap sorting in the steel industry.
Using camera images, the system is intended to recognise and classify different types of steel scrap. According to the consortium, the originally very large image-analysis model was significantly reduced through knowledge distillation and automated network optimisation. The result was a set of more compact models with lower energy demand, some of which even performed better than the original version.
This use case shows why efficient AI is not just a data-centre topic. In industrial environments, AI models must often work reliably, economically and sometimes close to production processes. Smaller models can make deployment easier because they reduce infrastructure requirements and may be easier to integrate into existing systems.
How data centres can plan AI workloads
ESCADE also analysed the energy efficiency of data centres. For this purpose, the project developed a tool that can forecast the energy consumption and operating costs of AI models. The aim is to help companies plan compute-intensive applications and make inefficient processes visible early.
According to doctoral researcher Hannah Stein, the tool can make it possible to schedule processes that require large amounts of computing power at times when electricity prices are low. For companies using AI at scale, such planning could become increasingly important. It links AI development to operational energy management and turns efficiency into a practical decision factor.
That fits a broader trend in which computing and data-centre infrastructure are becoming strategic elements of digital transformation. AI no longer depends only on algorithms and models. It also depends on how efficiently computing capacity, memory, storage, cooling and energy supply can be orchestrated.
Which architectures could lower consumption further
Beyond conventional model optimisation, ESCADE also investigated neuromorphic processors. These chips are inspired by the way biological nervous systems work and are designed to process information differently from classic computing architectures.
Initial tests suggested that such processors could already operate up to six times more efficiently than conventional architectures. For AI developers, this is relevant because future efficiency gains may not come only from software. They may also depend on new hardware concepts, from neuromorphic chips to specialised accelerators and more energy-conscious system architectures.
This also matters for automotive and industrial applications, where AI is increasingly expected to run under constraints. In many use cases, efficient AI processing is not only about sustainability, but also about heat, cost, packaging, latency and reliability.
Why QC-Train turns to quantum computing
While ESCADE has delivered concrete approaches to reducing the energy consumption of existing AI systems, the next technological step is now coming into focus. The QC-Train project, short for Quantum Computing Enhanced Training, started in July 2026 and will examine how quantum computing can be used for the training of large AI models.
The project partners are Infoteam Software AG, the Fraunhofer Institute for Integrated Circuits IIS and OptWare GmbH. DATEV and Schaeffler are involved as associated partners. The project is funded through Bavaria’s collaborative research programme for digitalisation, BayVFP, and is planned to run for three years.
The starting point is clear: the growing use of large language models and specialised AI networks is continuously increasing the demand for computing power, memory and energy. At the same time, regulatory expectations are also becoming more important. The EU AI Act, for example, calls for greater transparency around the energy consumption of AI systems and supports the development of corresponding efficiency standards.
QC-Train therefore aims to shift particularly compute-intensive mathematical operations to quantum processors in future. Daniel Scherer, head of the Quantum Compilation research group at Fraunhofer IIS, said the project is intended to move the most resource-intensive mathematical operations from traditional hardware currently in use to quantum processors.
The project team will develop different training methods for use on quantum computers, examine their scalability and assess in which scenarios even today’s quantum systems could already offer economic and ecological advantages. The goal is to advance quantum-assisted training approaches to the point where they can make a measurable contribution to more powerful and resource-efficient AI.
