Kategorie: Infrastructure

Benjamin Paaßen, Stefanie Go, Maximilian Mayer, Benjamin Kiesewetter, Anne Krüger, Jonas Leschke, Christian M. Stracke for the Research network Artificial Intelligence and Digital Autonomy in Research and Education (AIDARE).

Click here for the German version.

For future research and teaching in universities, access to large language models (LLMs) will be crucial. As such, universities ought to avoid dependencies on proprietary LLM suppliers and, instead, build a diversified AI infrastructure that supports rather than undermines digital autonomy of students, teachers, researchers, and the university as a whole. This document is directed at university leadership to support strategic steps toward such an infrastructure that can be implemented in the short- and mid-term with tangible benefits to digital autonomy.

Why digital autonomy?

With respect to artificial intelligence (AI), digital autonomy is core to the purpose of universities as teaching and research institutions: As institutions, they ought to be independent from the influence of AI hyperscalers; university researchers ought to be free to choose their own research tools and objectives instead of being limited by the constraints of proprietary AI systems; university teachers ought to be able to choose whether and how to integrate AI systems into their learning design and didactics without sending personal data of their students to servers abroad; and university students ought to become responsible citizens and autonomous experts in their field without offloading their cognition and academic responsibility to AI tools. Therefore, universities should build an AI infrastructure that promotes, rather than undermines digital autonomy in the sense of self-determination, epistemic agency and –justice, academic responsibility, and competencies (within and beyond specific academic subjects). Promoting digital autonomy in this ambitious sense involves many facets of universities, including teaching, research, administration, governance (such as guidelines), ethos. This document focuses on AI infrastructure, meaning the technological foundations for autonomy, such as algorithmic transparency and flexibility, that enable university members to manifest autonomy in the first place. We frame universities and their members as responsible actors that can shape the future of AI usage – instead of treating AI as an overwhelming external force – and we emphasize short- and mid-term steps universities can take to achieve tangible benefits to digital autonomy.

The status quo: LLM Chat-Interfaces

Many German universities have already made first steps toward more digital autonomy: They host their own website interfaces to chat with LLMs, such as HAWKI^[1] or KI:connect.nrw^[2]. These websites make sure that the account information of university members stay internal, while only the chat messages are forwarded to external providers who host the actual LLMs. This is a crucial first step for more data privacy and less dependency that can be implemented by universities at almost no additional cost (beyond what needs to be paid for LLM tokens, anyways). We recommend that universities open such interfaces for all their members, including students, teachers, researchers, and administrators to provide a meaningful alternative to proprietary systems. However, we emphasize that AI use must remain voluntary, and can even be discouraged in certain contexts (e.g., in teaching when building foundational knowledge and skills that are needed to competently judge AI outputs).

The next step: OpenWeight LLM Hosting

Just offering a chat interface is an insufficient basis for digital autonomy. If no further steps are taken, universities remain dependent on external providers of proprietary LLMs. The competitive position of these providers is further strengthened by the high-volume contracts with universities as well as the valuable research- and teaching-related chat data provided by universities – thus, potentially, deepening dependencies and leading to lock-in effects. Finally, privacy concerns remain as the chat messages themselves may leak personal data. Therefore, a diversified AI infrastructure is needed, meaning a diversity of hosters and a diversity of LLMs.

Some universities have, therefore, partnered with high performance computing (HPC) centers which host OpenWeight LLMs, such as Meta’s Llama models,  DeepSeek models, or even more open models, such as Apertus^[3]. Such arrangements have crucial advantages for universities: They can alleviate privacy concerns, can guarantee access to transparent models with known parameters, and can control costs more reliably. Germany already has working best-practice examples, most notably the GWDG^[4], which connects to dozens of universities, but other initiatives like Open Source-KI.nrw^[5] have started in this direction, as well. Universities should secure contracts with such OpenWeight LLM hosters to enable LLM access without privacy or dependency concerns for their members. If such hosters are not yet available, universities should partner with HPC institutions to enable OpenWeight LLM hosting on their servers. Such partnerships have also been recommended by the GWDG paper on AI basic infrastructure (“KI-Grundversorgung”)^[4]. To promote digital autonomy we recommend hosting at multiple HPCs, meaning more technical redundancy, LLM hosting capabilities in more locations, and less dependency on single providers.

Supporting research with LLM API access

A chatbot interface supports research processes at very small scales but is insufficient for larger research applications, such as automatically annotating/classifying large amounts of text data, transcription tasks, or building custom systems that need LLMs as a component. Research applications of LLMs are not limited to fields like computer science and computer linguistics but cross disciplinary boundaries, including natural sciences, social science and humanities. Establishing valid scientific methods with LLM involvement is an on-going process. Developing and applying such methods requires reliable LLM access with full transparency.

For such research use, an application programming interface (API) is required. Currently, API access to LLMs is (almost) only offered by proprietary vendors. However, depending on vendors who keep training data, LLM architecture and surrounding software secret, severely limits researchers’ epistemic agency in the sense of critically engaging with underlying biases in the LLMs, as well as good scientific practice in the sense of transparency and reproducibility of research. To build an alternative that promotes the digital autonomy of researchers, HPC centers must be equipped to offer APIs to researchers, which means handling a large expected volume of research-related LLM inference queries, far exceeding the 10 queries per day and user currently estimated^[4]. Importantly, this inference infrastructure is separate from and additional to classic scientific computing, also situated at HPCs: Inference APIs are intended for prototyping and small-compute research tasks, whereas classic scientific computing typically requires a proposal to apply for a large, multi-hour to multi-week computing effort. Classic scientific computing will still be needed, not least to fine-tune and train LLMs.

Universities should strategically apply for and politically demand investments in HPC centers to equip them with hardware and personnel to handle large volumes of research-related LLM inference queries and to provide API access to all researchers. This ensures that researchers have full freedom to choose whether and which LLM to use and can research every aspect of the models.

LLM integration in open source digital teaching tools

In teaching, chatbot interfaces are already useful for students and provide an alternative to personal accounts with proprietary vendors. However, many useful teaching applications require additional functionality, such as tutoring chatbots that should be able to answer questions and provide feedback and hints that are based on the material in one course. If such teaching applications are offered, they should not force students (or teachers) to transmit teaching material or student data to proprietary vendors; and teachers as well as students should have full autonomy how the prompts to LLMs are configured and which data is used to support teaching and learning. OpenSourceKI.nrw and GWDG have already developed prototype systems in this direction; the practice projects of KI:edu.nrw have shown how such infrastructure can be used in teaching^[6]. Universities should support open source developments that equip digital learning tools with freely configurable open weight LLM functionalities, including retrieval augmented generation (RAG) based on teaching material, and give their teachers and students the choice to integrate these functionalities in their courses.
We emphasize that we advocate for an autonomy-respecting option to use LLMs if desired. Universities should facilitate discussions for informed decisions by teachers and students regarding LLM use in specific learning contexts (i.e. a specific course in a specific subject for a specific learner).

The timeline

We believe that university chat interfaces and OpenWeight LLM hosting are steps that can be taken immediately or within months. To achieve API access for researchers and OpenSource teaching tools, prototypes already exists and universities should take steps to facilitate investments and developments (e.g., via proposals and political advocacy) and build partnerships with institutions (such as HPCs) who can become their suppliers for API access and OpenSource teaching tools. With coordinated effort, we believe that even these mid-term goals can be achieved within two years. We emphasize that these are only short- and mid-term steps to provide a technological foundation for digital autonomy at the university level. Universities will need to take additional steps in education, research, administration, governance, etc. Further, policy action on the national or even European level will be needed to achieve an autonomy-promoting infrastructure for training LLMs and gathering training data in a way that respects autonomy.

Related Initiatives

We are not the first to propose similar activities. The recommendations in this paper are particularly well aligned with the strategy paper of KI:edu.nrw^[7], the “KI-Zukunftsfonds Hochschule”^[8], the “KI Grundversorgung”^[4] and the expert hearings on “Souveräne KI-Infrastrukturen” of the Hochschulforum Digitalisierung in Germany. Other initiatives toward high performance computing hardware for AI are the AI (Giga-)factories (e.g.  HammerHAI^[9]), the JUPITER system^[10] at FZ Jülich, supercomputing for LLM training in Darmstadt^[11]. Related initiatives for the training of fully open LLMs in Europe are OpenEuroLLM^[12], the Swiss AI Iniative^[3] and Open GPT-X^[13]. In terms of promoting digital autonomy of all university members and remaining skeptical toward AI hype, our document aligns with the guidelines “Ethical AI in Higher Education” for teachers^[14] and for students^[15] (both by the Network “Ethical Use of AI”^[16]). All these initiatives (and many more) play a role to build an infrastructure that promotes digital autonomy in universities.

 

---

[1] https://hawki.hawk.de/ 

[2] https://kiconnect.pages.rwth-aachen.de/pages/ 

[3] https://ethz.ch/de/news-und-veranstaltungen/eth-news/news/2025/09/medienmitteilung-apertus-ein-vollstaendig-offenes-transparentes-und-mehrsprachiges-sprachmodell.html 

[4] https://kisski.gwdg.de/dok/grundversorgung.pdf 

[5] https://www.oski.nrw/ 

[6] https://ki-edu-nrw.ruhr-uni-bochum.de/ueber-das-projekt/phase-2/praxis-transferprojekte/aktuelle-praxisprojekte/ 

[7] https://ki-edu-nrw.ruhr-uni-bochum.de/wp-content/uploads/2025/07/2025_07_09_KI-Strategiepapier_NRW.pdf 

[8] https://www.stifterverband.org/sites/default/files/2025-02/ki-zukunftsfords_hochschulen_2026-2030.pdf 

[9] https://www.hlrs.de/press/detail/hammerhai-to-create-an-ai-factory-for-science-and-industry 

[10] https://www.fz-juelich.de/de/aktuelles/news/pressemitteilungen/2025/europas-ki-turbo-jupiter-ai-factory 

[11] https://hessian.ai/supercomputer-for-cutting-edge-ai-research-in-hesse/ 

[12] https://openeurollm.eu/ 

[13] https://opengpt-x.de/en/ 

[14] https://doi.org/10.5281/zenodo.10995669 (German version: https://doi.org/10.5281/zenodo.10793844)

[15] https://doi.org/10.5281/zenodo.15880726 

[16] https://ethischeki.ecompetence.eu 

Benjamin Paaßen, 2025-07-06

Large language models (LLMs) are a foundational technology, unlocking novel research methods, teaching practices, and business models – even when looking beyond the hype^[1]. Given the increasing importance of LLMs, it is deeply concerning that the supply chain for LLMs is controlled by a handful of AI corporations located in the US and China. The current practices of this handful of AI corporations stand in stark contrast to the vision of trustworthy AI, as well as human autonomy^[2]: their LLM-based bots spread misinformation and propaganda and are used to replace human labor; the AI platforms form an oligopoly that can dictate prices and conditions; and the data used for training has been gathered without consent. The alignment of current big AI players with autocratic regimes in China and the US only heightens the concern that AI tools will increasingly undermine, rather than strengthen, digital autonomy (consider the case of Microsoft cutting off services for ICC members). To maintain autonomy – as well as competitiveness for all companies that wish to remain independent of a tech oligopoly – alternatives along all steps of the LLM supply chain have to be established. In this paper, we focus on the software side of this supply chain, starting at the end users interacting with AI tools, over the deployment of LLMs for these tools, the training of such LLMs, to the training data for this training. Starting from the most urgent recommendations at the end user side, we provide recommendations to promote human autonomy at each step of this supply chain.

LLM-based Tools

End users most immediately engage with LLMs via tools, most notably chat interfaces such as ChatGPT. To support digital autonomy of end users, we therefore need to make sure that they do not become dependent on certain tools but have alternatives. This is particularly urgent since any delay will mean that end users will become locked into platforms and products that use the usage fees and the accumulated (person-related) information to strengthen their market position even further.

Hence, we need to offer alternatives for the most crucial tools, especially chat interfaces, research tools for scientific literature, as well as core educational tools such as AI plugins for digital learning platforms. Crucially, such tools should be hosted at universities themselves to avoid flows of person-related data to third parties and enable universities to design and adjust such tools to their research and teaching needs. Fortunately, this is achievable as the compute needs for the tools themselves are modest, as is demonstrated by many success stories of  universities hosting their own chat interface alternatives, e.g. via KI:connect.nrw^[3] and HAWKI^[4]. For literature search and AI plugins, developments are still in progress and urgently needed.

We recommend to:

• Provide project-based funding opportunities to develop new tools, both inside universities (e.g. via the Stiftung Innovation Hochschullehre) and beyond (e.g. via OpenSource development grants or ministry funding).
• Set up permanent development teams at the state or federal level which can maintain tools (e.g. as OpenSource output of project based funding) that have proven crucial and develop them further. These could be embedded at AI competency hubs, as suggested by the “KI-Zukunftsfonds Hochschule”^[5].
• Equip universities with sufficient funding for permanent staff which can introduce tools at the university level (e.g. for RAG), and provide support and guidance to researchers, teachers, students and administrators how to utilize these tools responsibly (i.e., enhance AI literacy).

LLM Deployment

To enable LLM-based tools, LLMs must be available in the first place. In particular, this means copies of trained LLMs being deployed on powerful GPU servers which can respond to queries with low delay (a few seconds). Such deployment services can be bought from commercial providers – but this would make all tools (and hence their users) dependent on the AI oligopoly, again. Therefore, we urgently need alternative LLM deployment options. However, to make LLM deployment efficient, we need some level of centralization to profit from scaling effects and pooled expertise. High performance computing centers are, hence, the prime actors to provide this service. We also know that such deployments are achievable as GWDG in Göttingen^[6] and OpenSourceKI.nrw^[7] already provide success stories for effective and efficient deployment.

In line with the notion of a “KI-Zukunftsfonds Hochschule”⁵, we recommend to:

• Provide substantial funds to equip Tier 2 High Performance Computing Centers with GPU server infrastructure to deploy multiple parallel copies of state-of-the-art open weight LLMs (with ca. 100 bio. Parameters).
• Provide Tier 2 High Performance Computing Centers withpermanent staff to operate this infrastructure, update the models as needed, and develop new APIs for tool development. For research and teaching, this will have to be funded by the state and federal level (e.g. via ministry funds). For private companies, parallel infrastructure may be set up as part of AI (Giga-)factories, such as HammerHAI^[8], and re-finance itself via contracts.

LLM Training

In order to deploy LLMs, they need to be trained, first. Fortunately, several alternatives for open weight LLMs are provided by private actors (e.g. Llama models out of the US, DeepSeek out of China, or Mistral models out of France) with substantial investment. When deploying such pre-trained models, no data or power flows to the model creators and, due to alternatives being available, we avoid dependencies on single creators. Hence, there is no urgent need to train alternative models. However, there is no guarantee that open weight models will be continuously provided by private actors and the training practices themselves do not consistently respect principles of openness and autonomy^[9]. Hence, we need to take steps to become capable to train LLMs, and to provide better training practices for LLMs without engaging in an “AI race”. Since building such capabilities is challenging and costly, we suggest to centralize this effort at the EU level. In more detail, we recommend to:

• Provide substantial funds to equip at least one Tier 1/Tier 0 High Performance Computing Center with sufficient GPU infrastructure to train state-of-the-art LLMs at the order of 100 bio. Parameters. The JUPITER system^[10] at FZ Jülich provides a good practice example in this regard.
• Set up at least one large-scale training project with ca. 200 mio. EUR of funding for ca. 200 researchers and developers over ca. 3 years to demonstrate that open models can be trained. Such large-scale projects should pool expertise and staff across university research teams as well as research institutes and companies that have experience in training LLMs at the 8 bio. parameter level (e.g. Darmstadt^[11]). The OpenEuroLLM^[12] and Open GPT-X^[13] initiatives may be starting points.

LLM Training Data Collection

Current LLM training operates on data that has been collected without consent, is strongly biased towards the US-based, male, white, internet-affine population and is badly curated, containing vast amounts of toxic or at least questionable data^[14]. It also becomes increasingly clear that LLM development is limited by the fact that no further reservoirs of publicly accessible, high-quality texts will become available – everything that is available has already been used^[15]. Hence, to provide a basis for autonomy-respecting training of LLMs in the future, we recommend to take first steps toward a long-term collection project for training data at the global level. More specifically, we recommend to:

• Set up a ten-year, long-term, global data collection project to gather high-quality, curated text data from sources that are currently under-represented. This data should be gathered with explicit, informed consent for LLM training, guaranteeing that the resulting LLMs will be available as a commons. The data collection should consider both direct data donations by individual authors as well as negotiations with publishers and other text-owning institutions. The Common Pile project^[16] may be a starting point.
• Set up a network of data stewards and curators who implement this project and are funded under it, involving public libraries and NGOs (e.g. Wikimedia) with experience on licensing and maintaining open data. These data stewards should also ensure long-term data maintenance under the FAIR principles and should ensure that the data is only available for LLM training under a public commons license to prevent privatization without consent.

Conclusion

We emphasize that all these recommendations can be implemented in parallel to gain sufficient speed. First success stories and examples driven forward by competent actors are already available at every step. The only thing needed is political action to make a public AI infrastructure happen and, thus, significantly strengthen human digital autonomy in the AI age.

---

[1] https://doi.org/10.1007/s10648-025-10020-8

[2] https://doi.org/10.1007/978-981-97-8638-1_7 

[3] https://kiconnect.pages.rwth-aachen.de/pages/

[4] https://hawki.hawk.de/

[5] https://www.stifterverband.org/sites/default/files/2025-02/ki-zukunftsfords_hochschulen_2026-2030.pdf 

[6] https://gwdg.de/en/services/application-services/ai-services/

[7] https://www.oski.nrw/

[8] https://www.hlrs.de/press/detail/hammerhai-to-create-an-ai-factory-for-science-and-industry

[9] https://doi.org/10.1145/3630106.365900

[10] https://www.fz-juelich.de/de/aktuelles/news/pressemitteilungen/2025/europas-ki-turbo-jupiter-ai-factory

[11] https://hessian.ai/supercomputer-for-cutting-edge-ai-research-in-hesse/

[12] https://openeurollm.eu/

[13] https://opengpt-x.de/en/

[14] https://knowingmachines.org/models-all-the-way

[15] https://epoch.ai/blog/will-we-run-out-of-data-limits-of-llm-scaling-based-on-human-generated-data

[16] https://blog.eleuther.ai/common-pile/