On March 13 and 14, seven ambitious teams met in the light-flooded premises of the Berlin Data Space to start the first stage of the SPRIND Challenge Composite Learning. Their mission: to reshape the future of artificial intelligence – decentralized, secure and scalable. Unlike the physical venue in Berlin’s Hackescher Markt, the digital data space in this challenge is decentralized, enabling AI training across multiple locations without centralized data sharing.

At the heart of the challenge is the decentralized training of AI models—leveraging distributed data across heterogeneous hardware. This is no small feat. Participants must develop a new framework that allows companies in Germany and Europe to efficiently and securely train large-scale AI models. In the first stage of the challenge, the models are to be trained with different hardware systems on at least three cloud platforms in parallel to create a robust foundation for the framework.

The event began with each team presenting its approach to decentralized AI. While some focused on abstract orchestration and bundling of existing computing servers, others worked on developing efficient accelerators and compilers for a wide range of, in some cases, entirely new hardware systems. Each contribution plays a role in shaping the overall framework, and although teams are competing for the best solution, their diverse approaches create significant synergy effects. The exchange of ideas extended well beyond the pitch presentations—over a shared dinner, participants had the opportunity to exchange knowledge and explore potential collaborations.

On the second day, specific technical challenges were deepened in workshop formats. In peer-to-peer sessions, the teams addressed topics such as hardware architectures, tools and datasets and discussed ways to attract the best talent to their companies. Special highlights were the discussions with external experts Arthur Douillard from Google DeepMind and Max Ryabinin from Together AI, who provided insights into current research results. One topic of discussion was the DiLoCo algorithm, which offers an innovative approach to decentralized optimization of model parameters. Instead of communicating all model parameters at the same time, this happens sequentially and only partially, which means that less data is moved at the same time and the individual computing nodes are given more autonomy. This significantly reduces the load on the infrastructure while maintaining or even increasing the quality of training – a decisive advance for decentralized AI applications.

Later in the day, Marco Schuldt from Germany’s Federal Ministry for Economic Affairs and Climate Action (BMWK) presented the IPCEI-CIS initiative—a key EU project aimed at strengthening Europe’s digital sovereignty. The ensuing discussion raised important questions: How can IPCEI-CIS partners leverage developments from the challenge? How can composite learning drive more efficient, flexible, and secure processes across industries?

The two days in Berlin have shown that the SPRIND Challenge is much more than a competition. It is a catalyst for innovation in the field of decentralized AI training. And perhaps a first step towards a European response to the global scaling race in AI development.

ABOUT THE CHALLENGE

Composite Learning combines distributed, decentralized and federated learning, offering a new approach to AI and allowing models to be trained across diverse systems without the need of centralized data centers. This method lets organizations collaborate and train models securely, preserving data privacy and making cutting-edge AI accessible to more companies. However, new solutions are needed to overcome the limitations of today’s systems, such as a lack of compatibility between different devices, communication bottlenecks, and reliance on central update servers.

The focus is on developing solutions that enable efficient model training on heterogeneous Hardware, from high-performance GPUs to CPUs of different types and manufacturers. Solutions must also be resilient, dynamically adapting to computing resource fluctuations and device outages. The teams will deliver a functional core for this framework as open source, which will serve as the foundation for further development, including commercial services and proprietary product features.

Seven teams started the challenge in February 2025 after their successful application. The first stage has a duration of twelve months and a maximum financing sum of up to 530,000 euros per team; the second lasts nine months and is financed with up to 520,000 euros. For the teams entering stage 3, SPRIND will provide up to 600,000 euros to promote further development and implementation.