AI Language Models: Germany Wants to Be In on the Game

ChatGPT amazes technology enthusiasts worldwide. The US multi-talent composes texts of all kinds and provides answers to complex questions. Companies in Germany and Europe also want to enter the market with new and intelligent AI applications. But the necessary open-source models, training data and sufficient computing infrastructures are lacking in this country. The feasibility study “Large AI Models for Germany” calculates what conditions are needed to make Germany a technology pioneer.

ChatGPT amazes technology enthusiasts worldwide. The US multi-talent composes texts of all kinds and provides answers to complex questions. Companies in Germany and Europe also want to enter the market with new and intelligent AI applications. But the necessary open-source models, training data and sufficient computing infrastructures are lacking in this country. The feasibility study “Large AI Models for Germany” calculates what conditions are needed to make Germany a technology pioneer.

The study, presented by the Large European AI Models (LEAM) initiative, shows that more than 80 percent of the experts surveyed recommend building an AI ecosystem and developing basic AI models based on European values. Currently, basic AI models that companies need to develop AI applications largely originate in the U.S. and China. This poses major challenges for European companies because, for example, access to proprietary services is more difficult and data is not processed in a DSGVO-compliant manner. “To ensure competitiveness and reduce dependencies, Germany needs its own basic AI models,” says Andreas Weiss, Head of Digital Business Models at eco – Association of the Internet Industry. To this end, the LEAM Initiative study now presents a concept for building a dedicated AI supercomputing infrastructure.

The need for language models is particularly great

Currently, 71 percent of local AI companies focus on large-scale language models. Multimodal AI models (38%) and AI models for computing business and manufacturing processes (34%) are also relevant to business. Basic AI models bring the next evolution in the artificial intelligence success story. Around two-thirds of German AI companies are already working with basic AI models. “In the next few years, even higher-performance models will come onto the market that are trained with larger data sets,” Weiss said.

Germany needs powerful AI computing infrastructure

A prerequisite for the development of AI foundation models is access to a powerful computing infrastructure. Building an AI data center is feasible in terms of hardware and software, according to the study’s authors, who include eco – Verband der Internetwirtschaft e.V.. They calculate that it would cost around 380 million euros to build and operate such an AI data center and have developed a model that includes public and private sector funding. Free open source models and available data for training the algorithms were named as further prerequisites by 58 percent of the experts surveyed.

The LEAM initiative advises integrating various values and principles in the development of basic AI models. These include data protection and data security, transparency, fairness and high data quality. This is to create trustworthy AI foundation models that offer broad application possibilities.

The feasibility study “Large AI Models for Germany” was commissioned by the Federal Ministry of Economics and Climate Protection (BMWK). The topic is now to be advanced in an exchange with politicians, industry and potential users. The LEAM initiative recommends the establishment of a project development company to build up the infrastructure and services.